((wrmask >> (i * 4)) & 0xf) << comp;
}
+static void
+set_writes_memory(const nir_shader *nir, struct radv_shader_info *info)
+{
+ if (nir->info.stage == MESA_SHADER_FRAGMENT)
+ info->ps.writes_memory = true;
+ else if (nir->info.stage == MESA_SHADER_GEOMETRY)
+ info->gs.writes_memory = true;
+}
+
static void
gather_intrinsic_store_deref_info(const nir_shader *nir,
const nir_intrinsic_instr *instr,
instr->intrinsic == nir_intrinsic_image_deref_atomic_xor ||
instr->intrinsic == nir_intrinsic_image_deref_atomic_exchange ||
instr->intrinsic == nir_intrinsic_image_deref_atomic_comp_swap) {
- if (nir->info.stage == MESA_SHADER_FRAGMENT)
- info->ps.writes_memory = true;
+ set_writes_memory(nir, info);
}
break;
}
case nir_intrinsic_ssbo_atomic_xor:
case nir_intrinsic_ssbo_atomic_exchange:
case nir_intrinsic_ssbo_atomic_comp_swap:
- if (nir->info.stage == MESA_SHADER_FRAGMENT)
- info->ps.writes_memory = true;
+ set_writes_memory(nir, info);
break;
case nir_intrinsic_load_deref:
gather_intrinsic_load_deref_info(nir, instr, info);
break;
case nir_intrinsic_store_deref:
gather_intrinsic_store_deref_info(nir, instr, info);
+ /* fallthrough */
+ case nir_intrinsic_deref_atomic_add:
+ case nir_intrinsic_deref_atomic_imin:
+ case nir_intrinsic_deref_atomic_umin:
+ case nir_intrinsic_deref_atomic_imax:
+ case nir_intrinsic_deref_atomic_umax:
+ case nir_intrinsic_deref_atomic_and:
+ case nir_intrinsic_deref_atomic_or:
+ case nir_intrinsic_deref_atomic_xor:
+ case nir_intrinsic_deref_atomic_exchange:
+ case nir_intrinsic_deref_atomic_comp_swap: {
+ if (nir_src_as_deref(instr->src[0])->mode & (nir_var_mem_global | nir_var_mem_ssbo))
+ set_writes_memory(nir, info);
break;
+ }
default:
break;
}
unsigned attrib_count = glsl_count_attribute_slots(var->type, true);
int idx = var->data.location;
- if (idx >= VERT_ATTRIB_GENERIC0 && idx <= VERT_ATTRIB_GENERIC15)
+ if (idx >= VERT_ATTRIB_GENERIC0 && idx < VERT_ATTRIB_GENERIC0 + MAX_VERTEX_ATTRIBS)
info->vs.has_vertex_buffers = true;
for (unsigned i = 0; i < attrib_count; ++i) {
case VARYING_SLOT_CLIP_DIST1:
info->ps.num_input_clips_culls += attrib_count;
break;
+ case VARYING_SLOT_VIEWPORT:
+ info->ps.viewport_index_input = true;
+ break;
default:
break;
}
if (var->data.interpolation == INTERP_MODE_FLAT)
info->ps.flat_shaded_mask |= mask << var->data.driver_location;
+ if (var->data.interpolation == INTERP_MODE_EXPLICIT)
+ info->ps.explicit_shaded_mask |= mask << var->data.driver_location;
if (var->data.location >= VARYING_SLOT_VAR0)
info->ps.input_mask |= mask << (var->data.location - VARYING_SLOT_VAR0);
default:
break;
}
+
+ if (idx >= FRAG_RESULT_DATA0 && idx <= FRAG_RESULT_DATA7) {
+ unsigned num_components = glsl_get_component_slots(glsl_without_array(var->type));
+ unsigned num_slots = glsl_count_attribute_slots(var->type, false);
+ unsigned write_mask = (1 << num_components) - 1;
+ unsigned slot = idx - FRAG_RESULT_DATA0;
+
+ for (unsigned i = 0; i < num_slots; i++) {
+ info->ps.cb_shader_mask |= write_mask << ((slot + i) * 4);
+ }
+ }
}
static void
radv_nir_shader_info_pass(const struct nir_shader *nir,
const struct radv_pipeline_layout *layout,
const struct radv_shader_variant_key *key,
- struct radv_shader_info *info)
+ struct radv_shader_info *info,
+ bool use_llvm)
{
struct nir_function *func =
(struct nir_function *)exec_list_get_head_const(&nir->functions);
info->loads_dynamic_offsets = true;
}
- nir_foreach_variable(variable, &nir->inputs)
+ nir_foreach_shader_in_variable(variable, nir)
gather_info_input_decl(nir, variable, info, key);
nir_foreach_block(block, func->impl) {
gather_info_block(nir, block, info);
}
- nir_foreach_variable(variable, &nir->outputs)
+ nir_foreach_shader_out_variable(variable, nir)
gather_info_output_decl(nir, variable, info, key);
if (nir->info.stage == MESA_SHADER_VERTEX ||
}
}
+ /* Make sure to export the ViewportIndex if the fragment shader needs it. */
+ if (key->vs_common_out.export_viewport_index) {
+ switch (nir->info.stage) {
+ case MESA_SHADER_VERTEX:
+ info->vs.output_usage_mask[VARYING_SLOT_VIEWPORT] |= 0x1;
+ break;
+ case MESA_SHADER_TESS_EVAL:
+ info->tes.output_usage_mask[VARYING_SLOT_VIEWPORT] |= 0x1;
+ break;
+ case MESA_SHADER_GEOMETRY:
+ info->gs.output_usage_mask[VARYING_SLOT_VIEWPORT] |= 0x1;
+ break;
+ default:
+ break;
+ }
+ }
+
if (nir->info.stage == MESA_SHADER_FRAGMENT)
info->ps.num_interp = nir->num_inputs;
info->ps.can_discard = nir->info.fs.uses_discard;
info->ps.early_fragment_test = nir->info.fs.early_fragment_tests;
info->ps.post_depth_coverage = nir->info.fs.post_depth_coverage;
+ info->ps.depth_layout = nir->info.fs.depth_layout;
break;
case MESA_SHADER_GEOMETRY:
info->gs.vertices_in = nir->info.gs.vertices_in;
info->tes.as_es = key->vs_common_out.as_es;
info->tes.export_prim_id = key->vs_common_out.export_prim_id;
info->is_ngg = key->vs_common_out.as_ngg;
+ info->is_ngg_passthrough = key->vs_common_out.as_ngg_passthrough;
break;
case MESA_SHADER_TESS_CTRL:
info->tcs.tcs_vertices_out = nir->info.tess.tcs_vertices_out;
info->vs.as_ls = key->vs_common_out.as_ls;
info->vs.export_prim_id = key->vs_common_out.export_prim_id;
info->is_ngg = key->vs_common_out.as_ngg;
+ info->is_ngg_passthrough = key->vs_common_out.as_ngg_passthrough;
break;
default:
break;
key->vs_common_out.as_es) {
struct radv_es_output_info *es_info =
nir->info.stage == MESA_SHADER_VERTEX ? &info->vs.es_info : &info->tes.es_info;
- uint32_t max_output_written = 0;
- uint64_t output_mask = nir->info.outputs_written;
- while (output_mask) {
- const int i = u_bit_scan64(&output_mask);
- unsigned param_index = shader_io_get_unique_index(i);
+ if (use_llvm) {
+ /* The outputs may contain gaps, use the highest output index + 1 */
+ uint32_t max_output_written = 0;
+ uint64_t output_mask = nir->info.outputs_written;
- max_output_written = MAX2(param_index, max_output_written);
- }
+ while (output_mask) {
+ const int i = u_bit_scan64(&output_mask);
+ unsigned param_index = shader_io_get_unique_index(i);
- es_info->esgs_itemsize = (max_output_written + 1) * 16;
+ max_output_written = MAX2(param_index, max_output_written);
+ }
+ es_info->esgs_itemsize = (max_output_written + 1) * 16;
+ } else {
+ /* The outputs don't contain gaps, se we can use the number of outputs */
+ uint32_t num_outputs_written = nir->info.stage == MESA_SHADER_VERTEX
+ ? info->vs.num_linked_outputs
+ : info->tes.num_linked_outputs;
+ es_info->esgs_itemsize = num_outputs_written * 16;
+ }
}
+ info->float_controls_mode = nir->info.float_controls_execution_mode;
+
+ if (nir->info.stage == MESA_SHADER_FRAGMENT) {
+ /* If the i-th output is used, all previous outputs must be
+ * non-zero to match the target format.
+ * TODO: compact MRT to avoid holes and to remove this
+ * workaround.
+ */
+ unsigned num_targets = (util_last_bit(info->ps.cb_shader_mask) + 3) / 4;
+ for (unsigned i = 0; i < num_targets; i++) {
+ if (!(info->ps.cb_shader_mask & (0xf << (i * 4)))) {
+ info->ps.cb_shader_mask |= 0xf << (i * 4);
+ }
+ }
+
+ if (key->fs.is_dual_src) {
+ info->ps.cb_shader_mask |= (info->ps.cb_shader_mask & 0xf) << 4;
+ }
+ }
}