#include "program.h"
+/**
+ * Get the varying type stripped of the outermost array if we're processing
+ * a stage whose varyings are arrays indexed by a vertex number (such as
+ * geometry shader inputs).
+ */
+static const glsl_type *
+get_varying_type(const ir_variable *var, gl_shader_stage stage)
+{
+ const glsl_type *type = var->type;
+
+ if (!var->data.patch &&
+ ((var->data.mode == ir_var_shader_out &&
+ stage == MESA_SHADER_TESS_CTRL) ||
+ (var->data.mode == ir_var_shader_in &&
+ (stage == MESA_SHADER_TESS_CTRL || stage == MESA_SHADER_TESS_EVAL ||
+ stage == MESA_SHADER_GEOMETRY)))) {
+ assert(type->is_array());
+ type = type->fields.array;
+ }
+
+ return type;
+}
+
/**
* Validate the types and qualifiers of an output from one stage against the
* matching input to another stage.
/* Check that the types match between stages.
*/
const glsl_type *type_to_match = input->type;
- if (consumer_stage == MESA_SHADER_GEOMETRY) {
- assert(type_to_match->is_array()); /* Enforced by ast_to_hir */
- type_to_match = type_to_match->element_type();
+
+ /* VS -> GS, VS -> TCS, VS -> TES, TES -> GS */
+ const bool extra_array_level = (producer_stage == MESA_SHADER_VERTEX &&
+ consumer_stage != MESA_SHADER_FRAGMENT) ||
+ consumer_stage == MESA_SHADER_GEOMETRY;
+ if (extra_array_level) {
+ assert(type_to_match->is_array());
+ type_to_match = type_to_match->fields.array;
}
+
if (type_to_match != output->type) {
/* There is a bit of a special case for gl_TexCoord. This
* built-in is unsized by default. Applications that variable
return;
}
+ if (input->data.patch != output->data.patch) {
+ linker_error(prog,
+ "%s shader output `%s' %s patch qualifier, "
+ "but %s shader input %s patch qualifier\n",
+ _mesa_shader_stage_to_string(producer_stage),
+ output->name,
+ (output->data.patch) ? "has" : "lacks",
+ _mesa_shader_stage_to_string(consumer_stage),
+ (input->data.patch) ? "has" : "lacks");
+ return;
+ }
+
if (!prog->IsES && input->data.invariant != output->data.invariant) {
linker_error(prog,
"%s shader output `%s' %s invariant qualifier, "
return;
}
- if (input->data.interpolation != output->data.interpolation) {
+ /* GLSL >= 4.40 removes text requiring interpolation qualifiers
+ * to match cross stage, they must only match within the same stage.
+ *
+ * From page 84 (page 90 of the PDF) of the GLSL 4.40 spec:
+ *
+ * "It is a link-time error if, within the same stage, the interpolation
+ * qualifiers of variables of the same name do not match.
+ *
+ */
+ if (input->data.interpolation != output->data.interpolation &&
+ prog->Version < 440) {
linker_error(prog,
"%s shader output `%s' specifies %s "
"interpolation qualifier, "
}
}
+/**
+ * Demote shader inputs and outputs that are not used in other stages, and
+ * remove them via dead code elimination.
+ */
+void
+remove_unused_shader_inputs_and_outputs(bool is_separate_shader_object,
+ gl_shader *sh,
+ enum ir_variable_mode mode)
+{
+ if (is_separate_shader_object)
+ return;
+
+ foreach_in_list(ir_instruction, node, sh->ir) {
+ ir_variable *const var = node->as_variable();
+
+ if ((var == NULL) || (var->data.mode != int(mode)))
+ continue;
+
+ /* A shader 'in' or 'out' variable is only really an input or output if
+ * its value is used by other shader stages. This will cause the
+ * variable to have a location assigned.
+ */
+ if (var->data.is_unmatched_generic_inout) {
+ assert(var->data.mode != ir_var_temporary);
+ var->data.mode = ir_var_auto;
+ }
+ }
+
+ /* Eliminate code that is now dead due to unused inputs/outputs being
+ * demoted.
+ */
+ while (do_dead_code(sh->ir, false))
+ ;
+
+}
/**
* Initialize this object based on a string that was passed to
this->location = -1;
this->orig_name = input;
- this->is_clip_distance_mesa = false;
+ this->lowered_builtin_array_variable = none;
this->skip_components = 0;
this->next_buffer_separator = false;
this->matched_candidate = NULL;
*/
if (ctx->Const.ShaderCompilerOptions[MESA_SHADER_VERTEX].LowerClipDistance &&
strcmp(this->var_name, "gl_ClipDistance") == 0) {
- this->is_clip_distance_mesa = true;
+ this->lowered_builtin_array_variable = clip_distance;
}
+
+ if (ctx->Const.LowerTessLevel &&
+ (strcmp(this->var_name, "gl_TessLevelOuter") == 0))
+ this->lowered_builtin_array_variable = tess_level_outer;
+ if (ctx->Const.LowerTessLevel &&
+ (strcmp(this->var_name, "gl_TessLevelInner") == 0))
+ this->lowered_builtin_array_variable = tess_level_inner;
}
this->matched_candidate->type->fields.array->matrix_columns;
const unsigned vector_elements =
this->matched_candidate->type->fields.array->vector_elements;
- unsigned actual_array_size = this->is_clip_distance_mesa ?
- prog->LastClipDistanceArraySize :
- this->matched_candidate->type->array_size();
+ const unsigned dmul =
+ this->matched_candidate->type->fields.array->is_double() ? 2 : 1;
+ unsigned actual_array_size;
+ switch (this->lowered_builtin_array_variable) {
+ case clip_distance:
+ actual_array_size = prog->LastClipDistanceArraySize;
+ break;
+ case tess_level_outer:
+ actual_array_size = 4;
+ break;
+ case tess_level_inner:
+ actual_array_size = 2;
+ break;
+ case none:
+ default:
+ actual_array_size = this->matched_candidate->type->array_size();
+ break;
+ }
if (this->is_subscripted) {
/* Check array bounds. */
actual_array_size);
return false;
}
- unsigned array_elem_size = this->is_clip_distance_mesa ?
- 1 : vector_elements * matrix_cols;
+ unsigned array_elem_size = this->lowered_builtin_array_variable ?
+ 1 : vector_elements * matrix_cols * dmul;
fine_location += array_elem_size * this->array_subscript;
this->size = 1;
} else {
}
this->vector_elements = vector_elements;
this->matrix_columns = matrix_cols;
- if (this->is_clip_distance_mesa)
+ if (this->lowered_builtin_array_variable)
this->type = GL_FLOAT;
else
this->type = this->matched_candidate->type->fields.array->gl_type;
if (!this->is_varying()) {
return 0;
}
-
return (this->num_components() + this->location_frac + 3)/4;
}
info->Outputs[info->NumOutputs].DstOffset = info->BufferStride[buffer];
++info->NumOutputs;
info->BufferStride[buffer] += output_size;
+ info->BufferStream[buffer] = this->stream_id;
num_components -= output_size;
location++;
location_frac = 0;
tfeedback_decl::find_candidate(gl_shader_program *prog,
hash_table *tfeedback_candidates)
{
- const char *name = this->is_clip_distance_mesa
- ? "gl_ClipDistanceMESA" : this->var_name;
+ const char *name = this->var_name;
+ switch (this->lowered_builtin_array_variable) {
+ case none:
+ name = this->var_name;
+ break;
+ case clip_distance:
+ name = "gl_ClipDistanceMESA";
+ break;
+ case tess_level_outer:
+ name = "gl_TessLevelOuterMESA";
+ break;
+ case tess_level_inner:
+ name = "gl_TessLevelInnerMESA";
+ break;
+ }
this->matched_candidate = (const tfeedback_candidate *)
hash_table_find(tfeedback_candidates, name);
if (!this->matched_candidate) {
class varying_matches
{
public:
- varying_matches(bool disable_varying_packing, bool consumer_is_fs);
+ varying_matches(bool disable_varying_packing,
+ gl_shader_stage producer_stage,
+ gl_shader_stage consumer_stage);
~varying_matches();
void record(ir_variable *producer_var, ir_variable *consumer_var);
- unsigned assign_locations();
+ unsigned assign_locations(struct gl_shader_program *prog,
+ uint64_t reserved_slots, bool separate_shader);
void store_locations() const;
private:
*/
unsigned matches_capacity;
- const bool consumer_is_fs;
+ gl_shader_stage producer_stage;
+ gl_shader_stage consumer_stage;
};
} /* anonymous namespace */
varying_matches::varying_matches(bool disable_varying_packing,
- bool consumer_is_fs)
+ gl_shader_stage producer_stage,
+ gl_shader_stage consumer_stage)
: disable_varying_packing(disable_varying_packing),
- consumer_is_fs(consumer_is_fs)
+ producer_stage(producer_stage),
+ consumer_stage(consumer_stage)
{
/* Note: this initial capacity is rather arbitrarily chosen to be large
* enough for many cases without wasting an unreasonable amount of space.
{
assert(producer_var != NULL || consumer_var != NULL);
- if ((producer_var && !producer_var->data.is_unmatched_generic_inout)
- || (consumer_var && !consumer_var->data.is_unmatched_generic_inout)) {
+ if ((producer_var && (!producer_var->data.is_unmatched_generic_inout ||
+ producer_var->data.explicit_location)) ||
+ (consumer_var && (!consumer_var->data.is_unmatched_generic_inout ||
+ consumer_var->data.explicit_location))) {
/* Either a location already exists for this variable (since it is part
* of fixed functionality), or it has already been recorded as part of a
* previous match.
}
if ((consumer_var == NULL && producer_var->type->contains_integer()) ||
- !consumer_is_fs) {
+ consumer_stage != MESA_SHADER_FRAGMENT) {
/* Since this varying is not being consumed by the fragment shader, its
* interpolation type varying cannot possibly affect rendering. Also,
* this variable is non-flat and is (or contains) an integer.
this->matches[this->num_matches].packing_order
= this->compute_packing_order(var);
if (this->disable_varying_packing) {
- unsigned slots = var->type->is_array()
- ? (var->type->length * var->type->fields.array->matrix_columns)
- : var->type->matrix_columns;
- this->matches[this->num_matches].num_components = 4 * slots;
+ unsigned slots;
+ gl_shader_stage stage =
+ (producer_var != NULL) ? producer_stage : consumer_stage;
+
+ const glsl_type *type = get_varying_type(var, stage);
+
+ slots = type->count_attribute_slots(false);
+ this->matches[this->num_matches].num_components = slots * 4;
} else {
this->matches[this->num_matches].num_components
= var->type->component_slots();
* passed to varying_matches::record().
*/
unsigned
-varying_matches::assign_locations()
+varying_matches::assign_locations(struct gl_shader_program *prog,
+ uint64_t reserved_slots,
+ bool separate_shader)
{
- /* Sort varying matches into an order that makes them easy to pack. */
- qsort(this->matches, this->num_matches, sizeof(*this->matches),
- &varying_matches::match_comparator);
+ /* We disable varying sorting for separate shader programs for the
+ * following reasons:
+ *
+ * 1/ All programs must sort the code in the same order to guarantee the
+ * interface matching. However varying_matches::record() will change the
+ * interpolation qualifier of some stages.
+ *
+ * 2/ GLSL version 4.50 removes the matching constrain on the interpolation
+ * qualifier.
+ *
+ * From Section 4.5 (Interpolation Qualifiers) of the GLSL 4.40 spec:
+ *
+ * "The type and presence of interpolation qualifiers of variables with
+ * the same name declared in all linked shaders for the same cross-stage
+ * interface must match, otherwise the link command will fail.
+ *
+ * When comparing an output from one stage to an input of a subsequent
+ * stage, the input and output don't match if their interpolation
+ * qualifiers (or lack thereof) are not the same."
+ *
+ * "It is a link-time error if, within the same stage, the interpolation
+ * qualifiers of variables of the same name do not match."
+ */
+ if (!separate_shader) {
+ /* Sort varying matches into an order that makes them easy to pack. */
+ qsort(this->matches, this->num_matches, sizeof(*this->matches),
+ &varying_matches::match_comparator);
+ }
unsigned generic_location = 0;
+ unsigned generic_patch_location = MAX_VARYING*4;
for (unsigned i = 0; i < this->num_matches; i++) {
+ unsigned *location = &generic_location;
+
+ const ir_variable *var;
+ const glsl_type *type;
+ bool is_vertex_input = false;
+ if (matches[i].consumer_var) {
+ var = matches[i].consumer_var;
+ type = get_varying_type(var, consumer_stage);
+ if (consumer_stage == MESA_SHADER_VERTEX)
+ is_vertex_input = true;
+ } else {
+ var = matches[i].producer_var;
+ type = get_varying_type(var, producer_stage);
+ }
+
+ if (var->data.patch)
+ location = &generic_patch_location;
+
/* Advance to the next slot if this varying has a different packing
* class than the previous one, and we're not already on a slot
* boundary.
if (i > 0 &&
this->matches[i - 1].packing_class
!= this->matches[i].packing_class) {
- generic_location = ALIGN(generic_location, 4);
+ *location = ALIGN(*location, 4);
+ }
+
+ unsigned num_elements = type->count_attribute_slots(is_vertex_input);
+ unsigned slot_end = this->disable_varying_packing ? 4 :
+ type->without_array()->vector_elements;
+ slot_end += *location - 1;
+
+ /* FIXME: We could be smarter in the below code and loop back over
+ * trying to fill any locations that we skipped because we couldn't pack
+ * the varying between an explicit location. For now just let the user
+ * hit the linking error if we run out of room and suggest they use
+ * explicit locations.
+ */
+ for (unsigned j = 0; j < num_elements; j++) {
+ while ((slot_end < MAX_VARYING * 4u) &&
+ ((reserved_slots & (UINT64_C(1) << *location / 4u) ||
+ (reserved_slots & (UINT64_C(1) << slot_end / 4u))))) {
+
+ *location = ALIGN(*location + 1, 4);
+ slot_end = *location;
+
+ /* reset the counter and try again */
+ j = 0;
+ }
+
+ /* Increase the slot to make sure there is enough room for next
+ * array element.
+ */
+ if (this->disable_varying_packing)
+ slot_end += 4;
+ else
+ slot_end += type->without_array()->vector_elements;
+ }
+
+ if (!var->data.patch && *location >= MAX_VARYING * 4u) {
+ linker_error(prog, "insufficient contiguous locations available for "
+ "%s it is possible an array or struct could not be "
+ "packed between varyings with explicit locations. Try "
+ "using an explicit location for arrays and structs.",
+ var->name);
}
- this->matches[i].generic_location = generic_location;
+ this->matches[i].generic_location = *location;
- generic_location += this->matches[i].num_components;
+ *location += this->matches[i].num_components;
}
return (generic_location + 3) / 4;
*
* Therefore, the packing class depends only on the interpolation type.
*/
- unsigned packing_class = var->data.centroid | (var->data.sample << 1);
+ unsigned packing_class = var->data.centroid | (var->data.sample << 1) |
+ (var->data.patch << 2);
packing_class *= 4;
packing_class += var->data.interpolation;
return packing_class;
populate_consumer_input_sets(void *mem_ctx, exec_list *ir,
hash_table *consumer_inputs,
hash_table *consumer_interface_inputs,
- ir_variable *consumer_inputs_with_locations[VARYING_SLOT_MAX])
+ ir_variable *consumer_inputs_with_locations[VARYING_SLOT_TESS_MAX])
{
memset(consumer_inputs_with_locations,
0,
- sizeof(consumer_inputs_with_locations[0]) * VARYING_SLOT_MAX);
+ sizeof(consumer_inputs_with_locations[0]) * VARYING_SLOT_TESS_MAX);
foreach_in_list(ir_instruction, node, ir) {
ir_variable *const input_var = node->as_variable();
const ir_variable *output_var,
hash_table *consumer_inputs,
hash_table *consumer_interface_inputs,
- ir_variable *consumer_inputs_with_locations[VARYING_SLOT_MAX])
+ ir_variable *consumer_inputs_with_locations[VARYING_SLOT_TESS_MAX])
{
ir_variable *input_var;
}
}
+/**
+ * Generate a bitfield map of the explicit locations for shader varyings.
+ *
+ * In theory a 32 bits value will be enough but a 64 bits value is future proof.
+ */
+uint64_t
+reserved_varying_slot(struct gl_shader *stage, ir_variable_mode io_mode)
+{
+ assert(io_mode == ir_var_shader_in || io_mode == ir_var_shader_out);
+ assert(MAX_VARYING <= 64); /* avoid an overflow of the returned value */
+
+ uint64_t slots = 0;
+ int var_slot;
+
+ if (!stage)
+ return slots;
+
+ foreach_in_list(ir_instruction, node, stage->ir) {
+ ir_variable *const var = node->as_variable();
+
+ if (var == NULL || var->data.mode != io_mode ||
+ !var->data.explicit_location ||
+ var->data.location < VARYING_SLOT_VAR0)
+ continue;
+
+ var_slot = var->data.location - VARYING_SLOT_VAR0;
+
+ unsigned num_elements = get_varying_type(var, stage->Stage)
+ ->count_attribute_slots(stage->Stage == MESA_SHADER_VERTEX);
+ for (unsigned i = 0; i < num_elements; i++) {
+ if (var_slot >= 0 && var_slot < MAX_VARYING)
+ slots |= UINT64_C(1) << var_slot;
+ var_slot += 1;
+ }
+ }
+
+ return slots;
+}
+
+
/**
* Assign locations for all variables that are produced in one pipeline stage
* (the "producer") and consumed in the next stage (the "consumer").
* each of these objects that matches one of the outputs of the
* producer.
*
- * \param gs_input_vertices: if \c consumer is a geometry shader, this is the
- * number of input vertices it accepts. Otherwise zero.
- *
* When num_tfeedback_decls is nonzero, it is permissible for the consumer to
* be NULL. In this case, varying locations are assigned solely based on the
* requirements of transform feedback.
struct gl_shader_program *prog,
gl_shader *producer, gl_shader *consumer,
unsigned num_tfeedback_decls,
- tfeedback_decl *tfeedback_decls,
- unsigned gs_input_vertices)
+ tfeedback_decl *tfeedback_decls)
{
- varying_matches matches(ctx->Const.DisableVaryingPacking,
- consumer && consumer->Stage == MESA_SHADER_FRAGMENT);
+ if (ctx->Const.DisableVaryingPacking) {
+ /* Transform feedback code assumes varyings are packed, so if the driver
+ * has disabled varying packing, make sure it does not support transform
+ * feedback.
+ */
+ assert(!ctx->Extensions.EXT_transform_feedback);
+ }
+
+ /* Tessellation shaders treat inputs and outputs as shared memory and can
+ * access inputs and outputs of other invocations.
+ * Therefore, they can't be lowered to temps easily (and definitely not
+ * efficiently).
+ */
+ bool disable_varying_packing =
+ ctx->Const.DisableVaryingPacking ||
+ (consumer && consumer->Stage == MESA_SHADER_TESS_EVAL) ||
+ (consumer && consumer->Stage == MESA_SHADER_TESS_CTRL) ||
+ (producer && producer->Stage == MESA_SHADER_TESS_CTRL);
+
+ varying_matches matches(disable_varying_packing,
+ producer ? producer->Stage : (gl_shader_stage)-1,
+ consumer ? consumer->Stage : (gl_shader_stage)-1);
hash_table *tfeedback_candidates
= hash_table_ctor(0, hash_table_string_hash, hash_table_string_compare);
hash_table *consumer_inputs
= hash_table_ctor(0, hash_table_string_hash, hash_table_string_compare);
hash_table *consumer_interface_inputs
= hash_table_ctor(0, hash_table_string_hash, hash_table_string_compare);
- ir_variable *consumer_inputs_with_locations[VARYING_SLOT_MAX] = {
+ ir_variable *consumer_inputs_with_locations[VARYING_SLOT_TESS_MAX] = {
NULL,
};
+ unsigned consumer_vertices = 0;
+ if (consumer && consumer->Stage == MESA_SHADER_GEOMETRY)
+ consumer_vertices = prog->Geom.VerticesIn;
+
/* Operate in a total of four passes.
*
* 1. Sort inputs / outputs into a canonical order. This is necessary so
/* If a matching input variable was found, add this ouptut (and the
* input) to the set. If this is a separable program and there is no
* consumer stage, add the output.
+ *
+ * Always add TCS outputs. They are shared by all invocations
+ * within a patch and can be used as shared memory.
*/
- if (input_var || (prog->SeparateShader && consumer == NULL)) {
+ if (input_var || (prog->SeparateShader && consumer == NULL) ||
+ producer->Type == GL_TESS_CONTROL_SHADER) {
matches.record(output_var, input_var);
}
matches.record(matched_candidate->toplevel_var, NULL);
}
- const unsigned slots_used = matches.assign_locations();
+ const uint64_t reserved_slots =
+ reserved_varying_slot(producer, ir_var_shader_out) |
+ reserved_varying_slot(consumer, ir_var_shader_in);
+
+ const unsigned slots_used = matches.assign_locations(prog, reserved_slots,
+ prog->SeparateShader);
matches.store_locations();
for (unsigned i = 0; i < num_tfeedback_decls; ++i) {
hash_table_dtor(consumer_inputs);
hash_table_dtor(consumer_interface_inputs);
- if (ctx->Const.DisableVaryingPacking) {
- /* Transform feedback code assumes varyings are packed, so if the driver
- * has disabled varying packing, make sure it does not support transform
- * feedback.
- */
- assert(!ctx->Extensions.EXT_transform_feedback);
- } else {
- if (producer) {
- lower_packed_varyings(mem_ctx, slots_used, ir_var_shader_out,
- 0, producer);
- }
- if (consumer) {
- lower_packed_varyings(mem_ctx, slots_used, ir_var_shader_in,
- gs_input_vertices, consumer);
- }
- }
-
if (consumer && producer) {
foreach_in_list(ir_instruction, node, consumer->ir) {
ir_variable *const var = node->as_variable();
var->name,
_mesa_shader_stage_to_string(producer->Stage));
}
-
- /* An 'in' variable is only really a shader input if its
- * value is written by the previous stage.
- */
- var->data.mode = ir_var_auto;
}
}
+
+ /* Now that validation is done its safe to remove unused varyings. As
+ * we have both a producer and consumer its safe to remove unused
+ * varyings even if the program is a SSO because the stages are being
+ * linked together i.e. we have a multi-stage SSO.
+ */
+ remove_unused_shader_inputs_and_outputs(false, producer,
+ ir_var_shader_out);
+ remove_unused_shader_inputs_and_outputs(false, consumer,
+ ir_var_shader_in);
+ }
+
+ if (!disable_varying_packing) {
+ if (producer) {
+ lower_packed_varyings(mem_ctx, slots_used, ir_var_shader_out,
+ 0, producer);
+ }
+ if (consumer) {
+ lower_packed_varyings(mem_ctx, slots_used, ir_var_shader_in,
+ consumer_vertices, consumer);
+ }
}
return true;
if (var && var->data.mode == ir_var_shader_out &&
var_counts_against_varying_limit(producer->Stage, var)) {
- output_vectors += var->type->count_attribute_slots();
+ /* outputs for fragment shader can't be doubles */
+ output_vectors += var->type->count_attribute_slots(false);
}
}
const unsigned output_components = output_vectors * 4;
if (output_components > max_output_components) {
if (ctx->API == API_OPENGLES2 || prog->IsES)
- linker_error(prog, "shader uses too many output vectors "
+ linker_error(prog, "%s shader uses too many output vectors "
"(%u > %u)\n",
+ _mesa_shader_stage_to_string(producer->Stage),
output_vectors,
max_output_components / 4);
else
- linker_error(prog, "shader uses too many output components "
+ linker_error(prog, "%s shader uses too many output components "
"(%u > %u)\n",
+ _mesa_shader_stage_to_string(producer->Stage),
output_components,
max_output_components);
if (var && var->data.mode == ir_var_shader_in &&
var_counts_against_varying_limit(consumer->Stage, var)) {
- input_vectors += var->type->count_attribute_slots();
+ /* vertex inputs aren't varying counted */
+ input_vectors += var->type->count_attribute_slots(false);
}
}
const unsigned input_components = input_vectors * 4;
if (input_components > max_input_components) {
if (ctx->API == API_OPENGLES2 || prog->IsES)
- linker_error(prog, "shader uses too many input vectors "
+ linker_error(prog, "%s shader uses too many input vectors "
"(%u > %u)\n",
+ _mesa_shader_stage_to_string(consumer->Stage),
input_vectors,
max_input_components / 4);
else
- linker_error(prog, "shader uses too many input components "
+ linker_error(prog, "%s shader uses too many input components "
"(%u > %u)\n",
+ _mesa_shader_stage_to_string(consumer->Stage),
input_components,
max_input_components);