void *mem_ctx;
int (*type_size)(const struct glsl_type *type);
nir_variable_mode modes;
+ nir_lower_io_options options;
};
void
nir_assign_var_locations(struct exec_list *var_list, unsigned *size,
- unsigned base_offset,
int (*type_size)(const struct glsl_type *))
{
unsigned location = 0;
- /* There are 32 regular and 32 patch varyings allowed */
- int locations[64][2];
- for (unsigned i = 0; i < 64; i++) {
- for (unsigned j = 0; j < 2; j++)
- locations[i][j] = -1;
- }
-
nir_foreach_variable(var, var_list) {
/*
* UBO's have their own address spaces, so don't count them towards the
var->interface_type != NULL)
continue;
- /* Make sure we give the same location to varyings packed with
- * ARB_enhanced_layouts.
- */
- int idx = var->data.location - base_offset;
- if (base_offset && idx >= 0) {
- assert(idx < ARRAY_SIZE(locations));
-
- if (locations[idx][var->data.index] == -1) {
- var->data.driver_location = location;
- locations[idx][var->data.index] = location;
- location += type_size(var->type);
- } else {
- var->data.driver_location = locations[idx][var->data.index];
- }
- } else {
- var->data.driver_location = location;
- location += type_size(var->type);
- }
+ var->data.driver_location = location;
+ location += type_size(var->type);
}
*size = location;
}
/**
- * Returns true if we're processing a stage whose inputs are arrays indexed
- * by a vertex number (such as geometry shader inputs).
+ * Return true if the given variable is a per-vertex input/output array.
+ * (such as geometry shader inputs).
*/
-static bool
-is_per_vertex_input(struct lower_io_state *state, nir_variable *var)
+bool
+nir_is_per_vertex_io(nir_variable *var, gl_shader_stage stage)
{
- gl_shader_stage stage = state->builder.shader->stage;
+ if (var->data.patch || !glsl_type_is_array(var->type))
+ return false;
- return var->data.mode == nir_var_shader_in && !var->data.patch &&
- (stage == MESA_SHADER_TESS_CTRL ||
- stage == MESA_SHADER_TESS_EVAL ||
- stage == MESA_SHADER_GEOMETRY);
-}
+ if (var->data.mode == nir_var_shader_in)
+ return stage == MESA_SHADER_GEOMETRY ||
+ stage == MESA_SHADER_TESS_CTRL ||
+ stage == MESA_SHADER_TESS_EVAL;
-static bool
-is_per_vertex_output(struct lower_io_state *state, nir_variable *var)
-{
- gl_shader_stage stage = state->builder.shader->stage;
- return var->data.mode == nir_var_shader_out && !var->data.patch &&
- stage == MESA_SHADER_TESS_CTRL;
+ if (var->data.mode == nir_var_shader_out)
+ return stage == MESA_SHADER_TESS_CTRL;
+
+ return false;
}
static nir_ssa_def *
get_io_offset(nir_builder *b, nir_deref_var *deref,
nir_ssa_def **vertex_index,
- int (*type_size)(const struct glsl_type *))
+ int (*type_size)(const struct glsl_type *),
+ unsigned *component)
{
nir_deref *tail = &deref->deref;
*/
if (vertex_index != NULL) {
tail = tail->child;
- assert(tail->deref_type == nir_deref_type_array);
nir_deref_array *deref_array = nir_deref_as_array(tail);
nir_ssa_def *vtx = nir_imm_int(b, deref_array->base_offset);
*vertex_index = vtx;
}
+ if (deref->var->data.compact) {
+ assert(tail->child->deref_type == nir_deref_type_array);
+ assert(glsl_type_is_scalar(glsl_without_array(deref->var->type)));
+ nir_deref_array *deref_array = nir_deref_as_array(tail->child);
+ /* We always lower indirect dereferences for "compact" array vars. */
+ assert(deref_array->deref_array_type == nir_deref_array_type_direct);
+
+ const unsigned total_offset = *component + deref_array->base_offset;
+ const unsigned slot_offset = total_offset / 4;
+ *component = total_offset % 4;
+ return nir_imm_int(b, type_size(glsl_vec4_type()) * slot_offset);
+ }
+
/* Just emit code and let constant-folding go to town */
nir_ssa_def *offset = nir_imm_int(b, 0);
static nir_intrinsic_instr *
lower_load(nir_intrinsic_instr *intrin, struct lower_io_state *state,
- nir_ssa_def *vertex_index, nir_ssa_def *offset)
+ nir_ssa_def *vertex_index, nir_ssa_def *offset,
+ unsigned component)
{
+ const nir_shader *nir = state->builder.shader;
nir_variable *var = intrin->variables[0]->var;
nir_variable_mode mode = var->data.mode;
+ nir_ssa_def *barycentric = NULL;
nir_intrinsic_op op;
switch (mode) {
case nir_var_shader_in:
- op = vertex_index ? nir_intrinsic_load_per_vertex_input :
- nir_intrinsic_load_input;
+ if (nir->stage == MESA_SHADER_FRAGMENT &&
+ nir->options->use_interpolated_input_intrinsics &&
+ var->data.interpolation != INTERP_MODE_FLAT) {
+ assert(vertex_index == NULL);
+
+ nir_intrinsic_op bary_op;
+ if (var->data.sample ||
+ (state->options & nir_lower_io_force_sample_interpolation))
+ bary_op = nir_intrinsic_load_barycentric_sample;
+ else if (var->data.centroid)
+ bary_op = nir_intrinsic_load_barycentric_centroid;
+ else
+ bary_op = nir_intrinsic_load_barycentric_pixel;
+
+ barycentric = nir_load_barycentric(&state->builder, bary_op,
+ var->data.interpolation);
+ op = nir_intrinsic_load_interpolated_input;
+ } else {
+ op = vertex_index ? nir_intrinsic_load_per_vertex_input :
+ nir_intrinsic_load_input;
+ }
break;
case nir_var_shader_out:
op = vertex_index ? nir_intrinsic_load_per_vertex_output :
nir_intrinsic_set_base(load, var->data.driver_location);
if (mode == nir_var_shader_in || mode == nir_var_shader_out)
- nir_intrinsic_set_component(load, var->data.location_frac);
+ nir_intrinsic_set_component(load, component);
if (load->intrinsic == nir_intrinsic_load_uniform)
nir_intrinsic_set_range(load, state->type_size(var->type));
- if (vertex_index)
+ if (vertex_index) {
load->src[0] = nir_src_for_ssa(vertex_index);
-
- load->src[vertex_index ? 1 : 0] = nir_src_for_ssa(offset);
+ load->src[1] = nir_src_for_ssa(offset);
+ } else if (barycentric) {
+ load->src[0] = nir_src_for_ssa(barycentric);
+ load->src[1] = nir_src_for_ssa(offset);
+ } else {
+ load->src[0] = nir_src_for_ssa(offset);
+ }
return load;
}
static nir_intrinsic_instr *
lower_store(nir_intrinsic_instr *intrin, struct lower_io_state *state,
- nir_ssa_def *vertex_index, nir_ssa_def *offset)
+ nir_ssa_def *vertex_index, nir_ssa_def *offset,
+ unsigned component)
{
nir_variable *var = intrin->variables[0]->var;
nir_variable_mode mode = var->data.mode;
nir_intrinsic_set_base(store, var->data.driver_location);
if (mode == nir_var_shader_out)
- nir_intrinsic_set_component(store, var->data.location_frac);
+ nir_intrinsic_set_component(store, component);
nir_intrinsic_set_write_mask(store, nir_intrinsic_write_mask(intrin));
nir_intrinsic_set_base(atomic, var->data.driver_location);
atomic->src[0] = nir_src_for_ssa(offset);
- for (unsigned i = 0; i < nir_op_infos[intrin->intrinsic].num_inputs; i++) {
+ for (unsigned i = 0; i < nir_intrinsic_infos[intrin->intrinsic].num_srcs; i++) {
nir_src_copy(&atomic->src[i+1], &intrin->src[i], atomic);
}
return atomic;
}
+static nir_intrinsic_instr *
+lower_interpolate_at(nir_intrinsic_instr *intrin, struct lower_io_state *state,
+ nir_ssa_def *offset, unsigned component)
+{
+ nir_variable *var = intrin->variables[0]->var;
+
+ assert(var->data.mode == nir_var_shader_in);
+
+ /* Ignore interpolateAt() for flat variables - flat is flat. */
+ if (var->data.interpolation == INTERP_MODE_FLAT)
+ return lower_load(intrin, state, NULL, offset, component);
+
+ nir_intrinsic_op bary_op;
+ switch (intrin->intrinsic) {
+ case nir_intrinsic_interp_var_at_centroid:
+ bary_op = (state->options & nir_lower_io_force_sample_interpolation) ?
+ nir_intrinsic_load_barycentric_sample :
+ nir_intrinsic_load_barycentric_centroid;
+ break;
+ case nir_intrinsic_interp_var_at_sample:
+ bary_op = nir_intrinsic_load_barycentric_at_sample;
+ break;
+ case nir_intrinsic_interp_var_at_offset:
+ bary_op = nir_intrinsic_load_barycentric_at_offset;
+ break;
+ default:
+ unreachable("Bogus interpolateAt() intrinsic.");
+ }
+
+ nir_intrinsic_instr *bary_setup =
+ nir_intrinsic_instr_create(state->mem_ctx, bary_op);
+
+ nir_ssa_dest_init(&bary_setup->instr, &bary_setup->dest, 2, 32, NULL);
+ nir_intrinsic_set_interp_mode(bary_setup, var->data.interpolation);
+
+ if (intrin->intrinsic != nir_intrinsic_interp_var_at_centroid)
+ nir_src_copy(&bary_setup->src[0], &intrin->src[0], bary_setup);
+
+ nir_builder_instr_insert(&state->builder, &bary_setup->instr);
+
+ nir_intrinsic_instr *load =
+ nir_intrinsic_instr_create(state->mem_ctx,
+ nir_intrinsic_load_interpolated_input);
+ load->num_components = intrin->num_components;
+
+ nir_intrinsic_set_base(load, var->data.driver_location);
+ nir_intrinsic_set_component(load, component);
+
+ load->src[0] = nir_src_for_ssa(&bary_setup->dest.ssa);
+ load->src[1] = nir_src_for_ssa(offset);
+
+ return load;
+}
+
static bool
nir_lower_io_block(nir_block *block,
struct lower_io_state *state)
{
nir_builder *b = &state->builder;
+ const nir_shader_compiler_options *options = b->shader->options;
nir_foreach_instr_safe(instr, block) {
if (instr->type != nir_instr_type_intrinsic)
case nir_intrinsic_var_atomic_comp_swap:
/* We can lower the io for this nir instrinsic */
break;
+ case nir_intrinsic_interp_var_at_centroid:
+ case nir_intrinsic_interp_var_at_sample:
+ case nir_intrinsic_interp_var_at_offset:
+ /* We can optionally lower these to load_interpolated_input */
+ if (options->use_interpolated_input_intrinsics)
+ break;
default:
/* We can't lower the io for this nir instrinsic, so skip it */
continue;
b->cursor = nir_before_instr(instr);
- const bool per_vertex =
- is_per_vertex_input(state, var) || is_per_vertex_output(state, var);
+ const bool per_vertex = nir_is_per_vertex_io(var, b->shader->stage);
nir_ssa_def *offset;
nir_ssa_def *vertex_index = NULL;
+ unsigned component_offset = var->data.location_frac;
offset = get_io_offset(b, intrin->variables[0],
per_vertex ? &vertex_index : NULL,
- state->type_size);
+ state->type_size, &component_offset);
nir_intrinsic_instr *replacement;
switch (intrin->intrinsic) {
case nir_intrinsic_load_var:
- replacement = lower_load(intrin, state, vertex_index, offset);
+ replacement = lower_load(intrin, state, vertex_index, offset,
+ component_offset);
break;
case nir_intrinsic_store_var:
- replacement = lower_store(intrin, state, vertex_index, offset);
+ replacement = lower_store(intrin, state, vertex_index, offset,
+ component_offset);
break;
case nir_intrinsic_var_atomic_add:
replacement = lower_atomic(intrin, state, offset);
break;
- default:
+ case nir_intrinsic_interp_var_at_centroid:
+ case nir_intrinsic_interp_var_at_sample:
+ case nir_intrinsic_interp_var_at_offset:
+ assert(vertex_index == NULL);
+ replacement = lower_interpolate_at(intrin, state, offset,
+ component_offset);
break;
+
+ default:
+ continue;
}
if (nir_intrinsic_infos[intrin->intrinsic].has_dest) {
if (intrin->dest.is_ssa) {
nir_ssa_dest_init(&replacement->instr, &replacement->dest,
- intrin->num_components,
+ intrin->dest.ssa.num_components,
intrin->dest.ssa.bit_size, NULL);
nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
nir_src_for_ssa(&replacement->dest.ssa));
static void
nir_lower_io_impl(nir_function_impl *impl,
nir_variable_mode modes,
- int (*type_size)(const struct glsl_type *))
+ int (*type_size)(const struct glsl_type *),
+ nir_lower_io_options options)
{
struct lower_io_state state;
state.mem_ctx = ralloc_parent(impl);
state.modes = modes;
state.type_size = type_size;
+ state.options = options;
nir_foreach_block(block, impl) {
nir_lower_io_block(block, &state);
void
nir_lower_io(nir_shader *shader, nir_variable_mode modes,
- int (*type_size)(const struct glsl_type *))
+ int (*type_size)(const struct glsl_type *),
+ nir_lower_io_options options)
{
nir_foreach_function(function, shader) {
- if (function->impl)
- nir_lower_io_impl(function->impl, modes, type_size);
+ if (function->impl) {
+ nir_lower_io_impl(function->impl, modes, type_size, options);
+ }
}
}
case nir_intrinsic_load_ssbo:
case nir_intrinsic_load_per_vertex_input:
case nir_intrinsic_load_per_vertex_output:
+ case nir_intrinsic_load_interpolated_input:
case nir_intrinsic_store_output:
return &instr->src[1];
case nir_intrinsic_store_ssbo:
projects / mesa.git / blobdiff