nir_builder builder;
void *mem_ctx;
int (*type_size)(const struct glsl_type *type);
- nir_variable_mode mode;
+ nir_variable_mode modes;
};
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;
- var->data.driver_location = location;
- location += type_size(var->type);
+ /* 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;
+
+ /* A dvec3 can be packed with a double we need special handling
+ * for this as we are packing across two locations.
+ */
+ if (glsl_get_base_type(var->type) == GLSL_TYPE_DOUBLE &&
+ glsl_get_vector_elements(var->type) == 3) {
+ /* Hack around type_size functions that expect vectors to be
+ * padded out to vec4. If a float type is the same size as a
+ * double then the type size is padded to vec4, otherwise
+ * set the offset to two doubles which offsets the location
+ * past the first two components in dvec3 which were stored at
+ * the previous location.
+ */
+ unsigned dsize = type_size(glsl_double_type());
+ unsigned offset =
+ dsize == type_size(glsl_float_type()) ? dsize : dsize * 2;
+
+ locations[idx + 1][var->data.index] = location + offset;
+ }
+
+ 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);
+ }
}
*size = location;
return offset;
}
-static nir_intrinsic_op
-load_op(struct lower_io_state *state,
- nir_variable_mode mode, bool per_vertex)
+static nir_intrinsic_instr *
+lower_load(nir_intrinsic_instr *intrin, struct lower_io_state *state,
+ nir_ssa_def *vertex_index, nir_ssa_def *offset)
{
+ 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 = per_vertex ? 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)
+ 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 = per_vertex ? nir_intrinsic_load_per_vertex_output :
- nir_intrinsic_load_output;
+ op = vertex_index ? nir_intrinsic_load_per_vertex_output :
+ nir_intrinsic_load_output;
break;
case nir_var_uniform:
op = nir_intrinsic_load_uniform;
break;
+ case nir_var_shared:
+ op = nir_intrinsic_load_shared;
+ break;
default:
unreachable("Unknown variable mode");
}
- return op;
+
+ nir_intrinsic_instr *load = nir_intrinsic_instr_create(state->mem_ctx, op);
+ load->num_components = intrin->num_components;
+
+ 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);
+
+ if (load->intrinsic == nir_intrinsic_load_uniform)
+ nir_intrinsic_set_range(load, state->type_size(var->type));
+
+ if (vertex_index) {
+ load->src[0] = nir_src_for_ssa(vertex_index);
+ 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 bool
-nir_lower_io_block(nir_block *block, void *void_state)
+static nir_intrinsic_instr *
+lower_store(nir_intrinsic_instr *intrin, struct lower_io_state *state,
+ nir_ssa_def *vertex_index, nir_ssa_def *offset)
{
- struct lower_io_state *state = void_state;
+ nir_variable *var = intrin->variables[0]->var;
+ nir_variable_mode mode = var->data.mode;
- nir_builder *b = &state->builder;
+ nir_intrinsic_op op;
+ if (mode == nir_var_shared) {
+ op = nir_intrinsic_store_shared;
+ } else {
+ assert(mode == nir_var_shader_out);
+ op = vertex_index ? nir_intrinsic_store_per_vertex_output :
+ nir_intrinsic_store_output;
+ }
- nir_foreach_instr_safe(block, instr) {
- if (instr->type != nir_instr_type_intrinsic)
- continue;
+ nir_intrinsic_instr *store = nir_intrinsic_instr_create(state->mem_ctx, op);
+ store->num_components = intrin->num_components;
- nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+ nir_src_copy(&store->src[0], &intrin->src[0], store);
- if (intrin->intrinsic != nir_intrinsic_load_var &&
- intrin->intrinsic != nir_intrinsic_store_var)
- continue;
+ nir_intrinsic_set_base(store, var->data.driver_location);
- nir_variable_mode mode = intrin->variables[0]->var->data.mode;
+ if (mode == nir_var_shader_out)
+ nir_intrinsic_set_component(store, var->data.location_frac);
- if (state->mode != nir_var_all && state->mode != mode)
- continue;
+ nir_intrinsic_set_write_mask(store, nir_intrinsic_write_mask(intrin));
- if (mode != nir_var_shader_in &&
- mode != nir_var_shader_out &&
- mode != nir_var_uniform)
- continue;
+ if (vertex_index)
+ store->src[1] = nir_src_for_ssa(vertex_index);
- b->cursor = nir_before_instr(instr);
+ store->src[vertex_index ? 2 : 1] = nir_src_for_ssa(offset);
- switch (intrin->intrinsic) {
- case nir_intrinsic_load_var: {
- bool per_vertex =
- is_per_vertex_input(state, intrin->variables[0]->var) ||
- is_per_vertex_output(state, intrin->variables[0]->var);
+ return store;
+}
- nir_ssa_def *offset;
- nir_ssa_def *vertex_index;
+static nir_intrinsic_instr *
+lower_atomic(nir_intrinsic_instr *intrin, struct lower_io_state *state,
+ nir_ssa_def *offset)
+{
+ nir_variable *var = intrin->variables[0]->var;
- offset = get_io_offset(b, intrin->variables[0],
- per_vertex ? &vertex_index : NULL,
- state->type_size);
+ assert(var->data.mode == nir_var_shared);
- nir_intrinsic_instr *load =
- nir_intrinsic_instr_create(state->mem_ctx,
- load_op(state, mode, per_vertex));
- load->num_components = intrin->num_components;
+ nir_intrinsic_op op;
+ switch (intrin->intrinsic) {
+#define OP(O) case nir_intrinsic_var_##O: op = nir_intrinsic_shared_##O; break;
+ OP(atomic_exchange)
+ OP(atomic_comp_swap)
+ OP(atomic_add)
+ OP(atomic_imin)
+ OP(atomic_umin)
+ OP(atomic_imax)
+ OP(atomic_umax)
+ OP(atomic_and)
+ OP(atomic_or)
+ OP(atomic_xor)
+#undef OP
+ default:
+ unreachable("Invalid atomic");
+ }
- load->const_index[0] =
- intrin->variables[0]->var->data.driver_location;
+ nir_intrinsic_instr *atomic =
+ nir_intrinsic_instr_create(state->mem_ctx, op);
- if (per_vertex)
- load->src[0] = nir_src_for_ssa(vertex_index);
+ nir_intrinsic_set_base(atomic, var->data.driver_location);
- load->src[per_vertex ? 1 : 0] = nir_src_for_ssa(offset);
+ atomic->src[0] = nir_src_for_ssa(offset);
+ for (unsigned i = 0; i < nir_op_infos[intrin->intrinsic].num_inputs; i++) {
+ nir_src_copy(&atomic->src[i+1], &intrin->src[i], atomic);
+ }
- if (intrin->dest.is_ssa) {
- nir_ssa_dest_init(&load->instr, &load->dest,
- intrin->num_components, NULL);
- nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
- nir_src_for_ssa(&load->dest.ssa));
- } else {
- nir_dest_copy(&load->dest, &intrin->dest, state->mem_ctx);
- }
+ return atomic;
+}
- nir_instr_insert_before(&intrin->instr, &load->instr);
- nir_instr_remove(&intrin->instr);
+static nir_intrinsic_instr *
+lower_interpolate_at(nir_intrinsic_instr *intrin, struct lower_io_state *state,
+ nir_ssa_def *offset)
+{
+ 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);
+
+ nir_intrinsic_op bary_op;
+ switch (intrin->intrinsic) {
+ case nir_intrinsic_interp_var_at_centroid:
+ bary_op = 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, var->data.location_frac);
+
+ 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)
+ continue;
+
+ nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+
+ switch (intrin->intrinsic) {
+ case nir_intrinsic_load_var:
+ case nir_intrinsic_store_var:
+ case nir_intrinsic_var_atomic_add:
+ case nir_intrinsic_var_atomic_imin:
+ case nir_intrinsic_var_atomic_umin:
+ case nir_intrinsic_var_atomic_imax:
+ case nir_intrinsic_var_atomic_umax:
+ case nir_intrinsic_var_atomic_and:
+ case nir_intrinsic_var_atomic_or:
+ case nir_intrinsic_var_atomic_xor:
+ case nir_intrinsic_var_atomic_exchange:
+ 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;
}
- case nir_intrinsic_store_var: {
- assert(mode == nir_var_shader_out);
+ nir_variable *var = intrin->variables[0]->var;
+ nir_variable_mode mode = var->data.mode;
- nir_ssa_def *offset;
- nir_ssa_def *vertex_index;
+ if ((state->modes & mode) == 0)
+ continue;
- bool per_vertex =
- is_per_vertex_output(state, intrin->variables[0]->var);
+ if (mode != nir_var_shader_in &&
+ mode != nir_var_shader_out &&
+ mode != nir_var_shared &&
+ mode != nir_var_uniform)
+ continue;
- offset = get_io_offset(b, intrin->variables[0],
- per_vertex ? &vertex_index : NULL,
- state->type_size);
+ b->cursor = nir_before_instr(instr);
- nir_intrinsic_op store_op =
- per_vertex ? nir_intrinsic_store_per_vertex_output :
- nir_intrinsic_store_output;
+ const bool per_vertex =
+ is_per_vertex_input(state, var) || is_per_vertex_output(state, var);
- nir_intrinsic_instr *store = nir_intrinsic_instr_create(state->mem_ctx,
- store_op);
- store->num_components = intrin->num_components;
+ nir_ssa_def *offset;
+ nir_ssa_def *vertex_index = NULL;
- nir_src_copy(&store->src[0], &intrin->src[0], store);
+ offset = get_io_offset(b, intrin->variables[0],
+ per_vertex ? &vertex_index : NULL,
+ state->type_size);
- store->const_index[0] =
- intrin->variables[0]->var->data.driver_location;
+ nir_intrinsic_instr *replacement;
- /* Copy the writemask */
- store->const_index[1] = intrin->const_index[0];
+ switch (intrin->intrinsic) {
+ case nir_intrinsic_load_var:
+ replacement = lower_load(intrin, state, vertex_index, offset);
+ break;
- if (per_vertex)
- store->src[1] = nir_src_for_ssa(vertex_index);
+ case nir_intrinsic_store_var:
+ replacement = lower_store(intrin, state, vertex_index, offset);
+ break;
- store->src[per_vertex ? 2 : 1] = nir_src_for_ssa(offset);
+ case nir_intrinsic_var_atomic_add:
+ case nir_intrinsic_var_atomic_imin:
+ case nir_intrinsic_var_atomic_umin:
+ case nir_intrinsic_var_atomic_imax:
+ case nir_intrinsic_var_atomic_umax:
+ case nir_intrinsic_var_atomic_and:
+ case nir_intrinsic_var_atomic_or:
+ case nir_intrinsic_var_atomic_xor:
+ case nir_intrinsic_var_atomic_exchange:
+ case nir_intrinsic_var_atomic_comp_swap:
+ assert(vertex_index == NULL);
+ replacement = lower_atomic(intrin, state, offset);
+ break;
- nir_instr_insert_before(&intrin->instr, &store->instr);
- nir_instr_remove(&intrin->instr);
+ 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);
break;
- }
default:
- break;
+ continue;
}
+
+ if (nir_intrinsic_infos[intrin->intrinsic].has_dest) {
+ if (intrin->dest.is_ssa) {
+ nir_ssa_dest_init(&replacement->instr, &replacement->dest,
+ 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));
+ } else {
+ nir_dest_copy(&replacement->dest, &intrin->dest, state->mem_ctx);
+ }
+ }
+
+ nir_instr_insert_before(&intrin->instr, &replacement->instr);
+ nir_instr_remove(&intrin->instr);
}
return true;
static void
nir_lower_io_impl(nir_function_impl *impl,
- nir_variable_mode mode,
+ nir_variable_mode modes,
int (*type_size)(const struct glsl_type *))
{
struct lower_io_state state;
nir_builder_init(&state.builder, impl);
state.mem_ctx = ralloc_parent(impl);
- state.mode = mode;
+ state.modes = modes;
state.type_size = type_size;
- nir_foreach_block(impl, nir_lower_io_block, &state);
+ nir_foreach_block(block, impl) {
+ nir_lower_io_block(block, &state);
+ }
nir_metadata_preserve(impl, nir_metadata_block_index |
nir_metadata_dominance);
}
void
-nir_lower_io(nir_shader *shader, nir_variable_mode mode,
+nir_lower_io(nir_shader *shader, nir_variable_mode modes,
int (*type_size)(const struct glsl_type *))
{
- nir_foreach_function(shader, function) {
+ nir_foreach_function(function, shader) {
if (function->impl)
- nir_lower_io_impl(function->impl, mode, type_size);
+ nir_lower_io_impl(function->impl, modes, type_size);
}
}
case nir_intrinsic_load_output:
case nir_intrinsic_load_uniform:
return &instr->src[0];
+ case nir_intrinsic_load_ubo:
+ 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:
case nir_intrinsic_store_per_vertex_output:
return &instr->src[2];
default:
projects / mesa.git / blobdiff