nir_shader *shader;
uint32_t constant_data_offset;
Block *block;
- bool *divergent_vals;
std::unique_ptr<Temp[]> allocated;
std::unordered_map<unsigned, std::array<Temp,NIR_MAX_VEC_COMPONENTS>> allocated_vec;
Stage stage; /* Stage */
/* GS inputs */
Temp gs_wave_id;
- /* gathered information */
- uint64_t input_masks[MESA_SHADER_COMPUTE];
- uint64_t output_masks[MESA_SHADER_COMPUTE];
-
/* VS output information */
bool export_clip_dists;
unsigned num_clip_distances;
* block instead. This is so that we can use any SGPR live-out of the side
* without the branch without creating a linear phi in the invert or merge block. */
bool
-sanitize_if(nir_function_impl *impl, bool *divergent, nir_if *nif)
+sanitize_if(nir_function_impl *impl, nir_if *nif)
{
//TODO: skip this if the condition is uniform and there are no divergent breaks/continues?
}
bool
-sanitize_cf_list(nir_function_impl *impl, bool *divergent, struct exec_list *cf_list)
+sanitize_cf_list(nir_function_impl *impl, struct exec_list *cf_list)
{
bool progress = false;
foreach_list_typed(nir_cf_node, cf_node, node, cf_list) {
break;
case nir_cf_node_if: {
nir_if *nif = nir_cf_node_as_if(cf_node);
- progress |= sanitize_cf_list(impl, divergent, &nif->then_list);
- progress |= sanitize_cf_list(impl, divergent, &nif->else_list);
- progress |= sanitize_if(impl, divergent, nif);
+ progress |= sanitize_cf_list(impl, &nif->then_list);
+ progress |= sanitize_cf_list(impl, &nif->else_list);
+ progress |= sanitize_if(impl, nif);
break;
}
case nir_cf_node_loop: {
nir_loop *loop = nir_cf_node_as_loop(cf_node);
- progress |= sanitize_cf_list(impl, divergent, &loop->body);
+ progress |= sanitize_cf_list(impl, &loop->body);
break;
}
case nir_cf_node_function:
unsigned lane_mask_size = ctx->program->lane_mask.size();
ctx->shader = shader;
- ctx->divergent_vals = nir_divergence_analysis(shader, nir_divergence_view_index_uniform);
+ nir_divergence_analysis(shader, nir_divergence_view_index_uniform);
/* sanitize control flow */
nir_metadata_require(impl, nir_metadata_dominance);
- sanitize_cf_list(impl, ctx->divergent_vals, &impl->body);
+ sanitize_cf_list(impl, &impl->body);
nir_metadata_preserve(impl, (nir_metadata)~nir_metadata_block_index);
/* we'll need this for isel */
case nir_op_b2f16:
case nir_op_b2f32:
case nir_op_mov:
- type = ctx->divergent_vals[alu_instr->dest.dest.ssa.index] ? RegType::vgpr : RegType::sgpr;
+ type = nir_dest_is_divergent(alu_instr->dest.dest) ? RegType::vgpr : RegType::sgpr;
break;
case nir_op_bcsel:
- type = ctx->divergent_vals[alu_instr->dest.dest.ssa.index] ? RegType::vgpr : RegType::sgpr;
+ type = nir_dest_is_divergent(alu_instr->dest.dest) ? RegType::vgpr : RegType::sgpr;
/* fallthrough */
default:
for (unsigned i = 0; i < nir_op_infos[alu_instr->op].num_inputs; i++) {
case nir_intrinsic_load_global:
case nir_intrinsic_vulkan_resource_index:
case nir_intrinsic_load_shared:
- type = ctx->divergent_vals[intrinsic->dest.ssa.index] ? RegType::vgpr : RegType::sgpr;
+ type = nir_dest_is_divergent(intrinsic->dest) ? RegType::vgpr : RegType::sgpr;
break;
case nir_intrinsic_load_view_index:
type = ctx->stage == fragment_fs ? RegType::vgpr : RegType::sgpr;
if (tex->dest.ssa.bit_size == 64)
size *= 2;
- if (tex->op == nir_texop_texture_samples)
- assert(!ctx->divergent_vals[tex->dest.ssa.index]);
- if (ctx->divergent_vals[tex->dest.ssa.index])
+ if (tex->op == nir_texop_texture_samples) {
+ assert(!tex->dest.ssa.divergent);
+ }
+ if (nir_dest_is_divergent(tex->dest))
allocated[tex->dest.ssa.index] = Temp(0, RegClass(RegType::vgpr, size));
else
allocated[tex->dest.ssa.index] = Temp(0, RegClass(RegType::sgpr, size));
break;
}
- if (ctx->divergent_vals[phi->dest.ssa.index]) {
+ if (nir_dest_is_divergent(phi->dest)) {
type = RegType::vgpr;
} else {
type = RegType::sgpr;
if (outinfo->writes_pointsize || outinfo->writes_viewport_index || outinfo->writes_layer)
pos_written |= 1 << 1;
- uint64_t mask = ctx->output_masks[nir->info.stage];
+ uint64_t mask = nir->info.outputs_written;
while (mask) {
int idx = u_bit_scan64(&mask);
if (idx >= VARYING_SLOT_VAR0 || idx == VARYING_SLOT_LAYER ||
}
nir_foreach_variable(variable, &nir->outputs)
{
- if (ctx->stage == vertex_geometry_gs)
- variable->data.driver_location = util_bitcount64(ctx->output_masks[nir->info.stage] & ((1ull << variable->data.location) - 1ull)) * 4;
- else if (ctx->stage == vertex_es ||
- ctx->stage == vertex_ls ||
- ctx->stage == vertex_tess_control_hs)
- // TODO: make this more compact
- variable->data.driver_location = shader_io_get_unique_index((gl_varying_slot) variable->data.location) * 4;
- else if (ctx->stage == vertex_vs || ctx->stage == ngg_vertex_gs)
+ if (ctx->stage == vertex_vs || ctx->stage == ngg_vertex_gs)
variable->data.driver_location = variable->data.location * 4;
- else
- unreachable("Unsupported VS stage");
assert(variable->data.location >= 0 && variable->data.location <= UINT8_MAX);
ctx->output_drv_loc_to_var_slot[MESA_SHADER_VERTEX][variable->data.driver_location / 4] = variable->data.location;
radv_vs_output_info *outinfo = &ctx->program->info->vs.outinfo;
setup_vs_output_info(ctx, nir, outinfo->export_prim_id,
ctx->options->key.vs_common_out.export_clip_dists, outinfo);
- } else if (ctx->stage == vertex_geometry_gs || ctx->stage == vertex_es) {
- /* TODO: radv_nir_shader_info_pass() already sets this but it's larger
- * than it needs to be in order to set it better, we have to improve
- * radv_nir_shader_info_pass() because gfx9_get_gs_info() uses
- * esgs_itemsize and has to be done before compilation
- */
- /* radv_es_output_info *outinfo = &ctx->program->info->vs.es_info;
- outinfo->esgs_itemsize = util_bitcount64(ctx->output_masks[nir->info.stage]) * 16u; */
} else if (ctx->stage == vertex_ls) {
- ctx->tcs_num_inputs = util_last_bit64(ctx->args->shader_info->vs.ls_outputs_written);
+ ctx->tcs_num_inputs = ctx->program->info->vs.num_linked_outputs;
}
if (ctx->stage == ngg_vertex_gs && ctx->args->options->key.vs_common_out.export_prim_id) {
void setup_gs_variables(isel_context *ctx, nir_shader *nir)
{
- if (ctx->stage == vertex_geometry_gs || ctx->stage == tess_eval_geometry_gs) {
- nir_foreach_variable(variable, &nir->inputs) {
- variable->data.driver_location = util_bitcount64(ctx->input_masks[nir->info.stage] & ((1ull << variable->data.location) - 1ull)) * 4;
- }
- } else if (ctx->stage == geometry_gs) {
- //TODO: make this more compact
- nir_foreach_variable(variable, &nir->inputs) {
- variable->data.driver_location = shader_io_get_unique_index((gl_varying_slot)variable->data.location) * 4;
- }
- } else {
- unreachable("Unsupported GS stage.");
- }
+ if (ctx->stage == vertex_geometry_gs || ctx->stage == tess_eval_geometry_gs)
+ ctx->program->config->lds_size = ctx->program->info->gs_ring_info.lds_size; /* Already in units of the alloc granularity */
nir_foreach_variable(variable, &nir->outputs) {
variable->data.driver_location = variable->data.location * 4;
ctx->stage == vertex_tess_control_hs &&
ctx->args->options->key.tcs.input_vertices == nir->info.tess.tcs_vertices_out;
- if (ctx->stage == tess_control_hs) {
- ctx->tcs_num_inputs = ctx->args->options->key.tcs.num_inputs;
- } else if (ctx->stage == vertex_tess_control_hs) {
- ctx->tcs_num_inputs = util_last_bit64(ctx->args->shader_info->vs.ls_outputs_written);
-
- if (ctx->tcs_in_out_eq) {
- ctx->tcs_temp_only_inputs = ~nir->info.tess.tcs_cross_invocation_inputs_read &
- ~nir->info.inputs_read_indirectly &
- nir->info.inputs_read;
- }
- } else {
- unreachable("Unsupported TCS shader stage");
+ if (ctx->tcs_in_out_eq) {
+ ctx->tcs_temp_only_inputs = ~nir->info.tess.tcs_cross_invocation_inputs_read &
+ ~nir->info.inputs_read_indirectly &
+ nir->info.inputs_read;
}
- ctx->tcs_num_outputs = util_last_bit64(ctx->args->shader_info->tcs.outputs_written);
- ctx->tcs_num_patch_outputs = util_last_bit64(ctx->args->shader_info->tcs.patch_outputs_written);
+ ctx->tcs_num_inputs = ctx->program->info->tcs.num_linked_inputs;
+ ctx->tcs_num_outputs = ctx->program->info->tcs.num_linked_outputs;
+ ctx->tcs_num_patch_outputs = ctx->program->info->tcs.num_linked_patch_outputs;
ctx->tcs_num_patches = get_tcs_num_patches(
ctx->args->options->key.tcs.input_vertices,
void
setup_tcs_variables(isel_context *ctx, nir_shader *nir)
{
- nir_foreach_variable(variable, &nir->inputs) {
- variable->data.driver_location = shader_io_get_unique_index((gl_varying_slot) variable->data.location) * 4;
- }
-
nir_foreach_variable(variable, &nir->outputs) {
- variable->data.driver_location = shader_io_get_unique_index((gl_varying_slot) variable->data.location) * 4;
assert(variable->data.location >= 0 && variable->data.location <= UINT8_MAX);
+ if (variable->data.location == VARYING_SLOT_TESS_LEVEL_OUTER)
+ ctx->tcs_tess_lvl_out_loc = variable->data.driver_location * 4u;
+ else if (variable->data.location == VARYING_SLOT_TESS_LEVEL_INNER)
+ ctx->tcs_tess_lvl_in_loc = variable->data.driver_location * 4u;
+
if (variable->data.patch)
ctx->output_tcs_patch_drv_loc_to_var_slot[variable->data.driver_location / 4] = variable->data.location;
else
ctx->output_drv_loc_to_var_slot[MESA_SHADER_TESS_CTRL][variable->data.driver_location / 4] = variable->data.location;
}
-
- ctx->tcs_tess_lvl_out_loc = shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER) * 16u;
- ctx->tcs_tess_lvl_in_loc = shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER) * 16u;
}
void
setup_tes_variables(isel_context *ctx, nir_shader *nir)
{
ctx->tcs_num_patches = ctx->args->options->key.tes.num_patches;
- ctx->tcs_num_outputs = ctx->args->options->key.tes.tcs_num_outputs;
-
- nir_foreach_variable(variable, &nir->inputs) {
- variable->data.driver_location = shader_io_get_unique_index((gl_varying_slot) variable->data.location) * 4;
- }
+ ctx->tcs_num_outputs = ctx->program->info->tes.num_linked_inputs;
nir_foreach_variable(variable, &nir->outputs) {
if (ctx->stage == tess_eval_vs || ctx->stage == ngg_tess_eval_gs)
variable->data.driver_location = variable->data.location * 4;
- else if (ctx->stage == tess_eval_es)
- variable->data.driver_location = shader_io_get_unique_index((gl_varying_slot) variable->data.location) * 4;
- else if (ctx->stage == tess_eval_geometry_gs)
- variable->data.driver_location = util_bitcount64(ctx->output_masks[nir->info.stage] & ((1ull << variable->data.location) - 1ull)) * 4;
- else
- unreachable("Unsupported TES shader stage");
}
if (ctx->stage == tess_eval_vs || ctx->stage == ngg_tess_eval_gs) {
}
}
-void
-get_io_masks(isel_context *ctx, unsigned shader_count, struct nir_shader *const *shaders)
-{
- for (unsigned i = 0; i < shader_count; i++) {
- nir_shader *nir = shaders[i];
- if (nir->info.stage == MESA_SHADER_COMPUTE)
- continue;
-
- uint64_t output_mask = 0;
- nir_foreach_variable(variable, &nir->outputs) {
- const glsl_type *type = variable->type;
- if (nir_is_per_vertex_io(variable, nir->info.stage))
- type = type->fields.array;
- unsigned slots = type->count_attribute_slots(false);
- if (variable->data.compact) {
- unsigned component_count = variable->data.location_frac + type->length;
- slots = (component_count + 3) / 4;
- }
- output_mask |= ((1ull << slots) - 1) << variable->data.location;
- }
-
- uint64_t input_mask = 0;
- nir_foreach_variable(variable, &nir->inputs) {
- const glsl_type *type = variable->type;
- if (nir_is_per_vertex_io(variable, nir->info.stage))
- type = type->fields.array;
- unsigned slots = type->count_attribute_slots(false);
- if (variable->data.compact) {
- unsigned component_count = variable->data.location_frac + type->length;
- slots = (component_count + 3) / 4;
- }
- input_mask |= ((1ull << slots) - 1) << variable->data.location;
- }
-
- ctx->output_masks[nir->info.stage] |= output_mask;
- if (i + 1 < shader_count)
- ctx->input_masks[shaders[i + 1]->info.stage] |= output_mask;
-
- ctx->input_masks[nir->info.stage] |= input_mask;
- if (i)
- ctx->output_masks[shaders[i - 1]->info.stage] |= input_mask;
- }
-}
-
unsigned
lower_bit_size_callback(const nir_alu_instr *alu, void *_)
{
bool lower_to_scalar = false;
bool lower_pack = false;
+ nir_variable_mode robust_modes = (nir_variable_mode)0;
+
+ if (ctx->options->robust_buffer_access) {
+ robust_modes = (nir_variable_mode)(nir_var_mem_ubo |
+ nir_var_mem_ssbo |
+ nir_var_mem_global |
+ nir_var_mem_push_const);
+ }
+
if (nir_opt_load_store_vectorize(nir,
(nir_variable_mode)(nir_var_mem_ssbo | nir_var_mem_ubo |
nir_var_mem_push_const | nir_var_mem_shared |
nir_var_mem_global),
- mem_vectorize_callback)) {
+ mem_vectorize_callback, robust_modes)) {
lower_to_scalar = true;
lower_pack = true;
}
program->vgpr_limit = get_addr_vgpr_from_waves(program, program->min_waves);
program->sgpr_limit = get_addr_sgpr_from_waves(program, program->min_waves);
- get_io_masks(&ctx, shader_count, shaders);
-
unsigned scratch_size = 0;
if (program->stage == gs_copy_vs) {
assert(shader_count == 1);