nir/gather_info: Handle multi-slot variables in io bitfields
[mesa.git] / src / glsl / link_uniform_blocks.cpp
index c9cbde95b6f6ae7455fec5d6cebe4410742ba147..d5d30bb0a0dd39497084fc42e65a8865197e5099 100644 (file)
 #include "ir.h"
 #include "linker.h"
 #include "ir_uniform.h"
+#include "link_uniform_block_active_visitor.h"
+#include "util/hash_table.h"
 #include "program.h"
 
+namespace {
+
+class ubo_visitor : public program_resource_visitor {
+public:
+   ubo_visitor(void *mem_ctx, gl_uniform_buffer_variable *variables,
+               unsigned num_variables)
+      : index(0), offset(0), buffer_size(0), variables(variables),
+        num_variables(num_variables), mem_ctx(mem_ctx), is_array_instance(false)
+   {
+      /* empty */
+   }
+
+   void process(const glsl_type *type, const char *name)
+   {
+      this->offset = 0;
+      this->buffer_size = 0;
+      this->is_array_instance = strchr(name, ']') != NULL;
+      this->program_resource_visitor::process(type, name);
+   }
+
+   unsigned index;
+   unsigned offset;
+   unsigned buffer_size;
+   gl_uniform_buffer_variable *variables;
+   unsigned num_variables;
+   void *mem_ctx;
+   bool is_array_instance;
+
+private:
+   virtual void visit_field(const glsl_type *type, const char *name,
+                            bool row_major)
+   {
+      (void) type;
+      (void) name;
+      (void) row_major;
+      assert(!"Should not get here.");
+   }
+
+   virtual void enter_record(const glsl_type *type, const char *,
+                             bool row_major, const unsigned packing) {
+      assert(type->is_record());
+      if (packing == GLSL_INTERFACE_PACKING_STD430)
+         this->offset = glsl_align(
+            this->offset, type->std430_base_alignment(row_major));
+      else
+         this->offset = glsl_align(
+            this->offset, type->std140_base_alignment(row_major));
+   }
+
+   virtual void leave_record(const glsl_type *type, const char *,
+                             bool row_major, const unsigned packing) {
+      assert(type->is_record());
+
+      /* If this is the last field of a structure, apply rule #9.  The
+       * GL_ARB_uniform_buffer_object spec says:
+       *
+       *     "The structure may have padding at the end; the base offset of
+       *     the member following the sub-structure is rounded up to the next
+       *     multiple of the base alignment of the structure."
+       */
+      if (packing == GLSL_INTERFACE_PACKING_STD430)
+         this->offset = glsl_align(
+            this->offset, type->std430_base_alignment(row_major));
+      else
+         this->offset = glsl_align(
+            this->offset, type->std140_base_alignment(row_major));
+   }
+
+   virtual void visit_field(const glsl_type *type, const char *name,
+                            bool row_major, const glsl_type *,
+                            const unsigned packing,
+                            bool last_field)
+   {
+      assert(this->index < this->num_variables);
+
+      gl_uniform_buffer_variable *v = &this->variables[this->index++];
+
+      v->Name = ralloc_strdup(mem_ctx, name);
+      v->Type = type;
+      v->RowMajor = type->without_array()->is_matrix() && row_major;
+
+      if (this->is_array_instance) {
+         v->IndexName = ralloc_strdup(mem_ctx, name);
+
+         char *open_bracket = strchr(v->IndexName, '[');
+         assert(open_bracket != NULL);
+
+         char *close_bracket = strchr(open_bracket, '.') - 1;
+         assert(close_bracket != NULL);
+
+         /* Length of the tail without the ']' but with the NUL.
+          */
+         unsigned len = strlen(close_bracket + 1) + 1;
+
+         memmove(open_bracket, close_bracket + 1, len);
+      } else {
+         v->IndexName = v->Name;
+      }
+
+      unsigned alignment = 0;
+      unsigned size = 0;
+
+      /* From ARB_program_interface_query:
+       *
+       *     "If the final member of an active shader storage block is array
+       *      with no declared size, the minimum buffer size is computed
+       *      assuming the array was declared as an array with one element."
+       *
+       * For that reason, we use the base type of the unsized array to calculate
+       * its size. We don't need to check if the unsized array is the last member
+       * of a shader storage block (that check was already done by the parser).
+       */
+      const glsl_type *type_for_size = type;
+      if (type->is_unsized_array()) {
+         assert(last_field);
+         type_for_size = type->without_array();
+      }
+
+      if (packing == GLSL_INTERFACE_PACKING_STD430) {
+         alignment = type->std430_base_alignment(v->RowMajor);
+         size = type_for_size->std430_size(v->RowMajor);
+      } else {
+         alignment = type->std140_base_alignment(v->RowMajor);
+         size = type_for_size->std140_size(v->RowMajor);
+      }
+
+      this->offset = glsl_align(this->offset, alignment);
+      v->Offset = this->offset;
+
+      this->offset += size;
+
+      /* From the GL_ARB_uniform_buffer_object spec:
+       *
+       *     "For uniform blocks laid out according to [std140] rules, the
+       *      minimum buffer object size returned by the
+       *      UNIFORM_BLOCK_DATA_SIZE query is derived by taking the offset of
+       *      the last basic machine unit consumed by the last uniform of the
+       *      uniform block (including any end-of-array or end-of-structure
+       *      padding), adding one, and rounding up to the next multiple of
+       *      the base alignment required for a vec4."
+       */
+      this->buffer_size = glsl_align(this->offset, 16);
+   }
+};
+
+class count_block_size : public program_resource_visitor {
+public:
+   count_block_size() : num_active_uniforms(0)
+   {
+      /* empty */
+   }
+
+   unsigned num_active_uniforms;
+
+private:
+   virtual void visit_field(const glsl_type *type, const char *name,
+                            bool row_major)
+   {
+      (void) type;
+      (void) name;
+      (void) row_major;
+      this->num_active_uniforms++;
+   }
+};
+
+} /* anonymous namespace */
+
+struct block {
+   const glsl_type *type;
+   bool has_instance_name;
+};
+
+static void
+process_block_array(struct uniform_block_array_elements *ub_array, char **name,
+                    size_t name_length, gl_uniform_block *blocks,
+                    ubo_visitor *parcel, gl_uniform_buffer_variable *variables,
+                    const struct link_uniform_block_active *const b,
+                    unsigned *block_index, unsigned *binding_offset,
+                    struct gl_context *ctx, struct gl_shader_program *prog)
+{
+   if (ub_array) {
+      for (unsigned j = 0; j < ub_array->num_array_elements; j++) {
+        size_t new_length = name_length;
+
+         /* Append the subscript to the current variable name */
+         ralloc_asprintf_rewrite_tail(name, &new_length, "[%u]",
+                                      ub_array->array_elements[j]);
+
+         process_block_array(ub_array->array, name, new_length, blocks,
+                             parcel, variables, b, block_index,
+                             binding_offset, ctx, prog);
+      }
+   } else {
+      unsigned i = *block_index;
+      const glsl_type *type =  b->type->without_array();
+
+      blocks[i].Name = ralloc_strdup(blocks, *name);
+      blocks[i].Uniforms = &variables[(*parcel).index];
+
+      /* The GL_ARB_shading_language_420pack spec says:
+       *
+       *     "If the binding identifier is used with a uniform block
+       *     instanced as an array then the first element of the array
+       *     takes the specified block binding and each subsequent
+       *     element takes the next consecutive uniform block binding
+       *     point."
+       */
+      blocks[i].Binding = (b->has_binding) ? b->binding + *binding_offset : 0;
+
+      blocks[i].UniformBufferSize = 0;
+      blocks[i]._Packing = gl_uniform_block_packing(type->interface_packing);
+
+      parcel->process(type, blocks[i].Name);
+
+      blocks[i].UniformBufferSize = parcel->buffer_size;
+
+      /* Check SSBO size is lower than maximum supported size for SSBO */
+      if (b->is_shader_storage &&
+          parcel->buffer_size > ctx->Const.MaxShaderStorageBlockSize) {
+         linker_error(prog, "shader storage block `%s' has size %d, "
+                      "which is larger than than the maximum allowed (%d)",
+                      b->type->name,
+                      parcel->buffer_size,
+                      ctx->Const.MaxShaderStorageBlockSize);
+      }
+      blocks[i].NumUniforms =
+         (unsigned)(ptrdiff_t)(&variables[parcel->index] - blocks[i].Uniforms);
+      blocks[i].IsShaderStorage = b->is_shader_storage;
+
+      *block_index = *block_index + 1;
+      *binding_offset = *binding_offset + 1;
+   }
+}
+
+/* This function resizes the array types of the block so that later we can use
+ * this new size to correctly calculate the offest for indirect indexing.
+ */
+const glsl_type *
+resize_block_array(const glsl_type *type,
+                   struct uniform_block_array_elements *ub_array)
+{
+   if (type->is_array()) {
+      struct uniform_block_array_elements *child_array =
+         type->fields.array->is_array() ? ub_array->array : NULL;
+      const glsl_type *new_child_type =
+         resize_block_array(type->fields.array, child_array);
+
+      const glsl_type *new_type =
+         glsl_type::get_array_instance(new_child_type,
+                                       ub_array->num_array_elements);
+      ub_array->ir->array->type = new_type;
+      return new_type;
+   } else {
+      return type;
+   }
+}
+
+unsigned
+link_uniform_blocks(void *mem_ctx,
+                    struct gl_context *ctx,
+                    struct gl_shader_program *prog,
+                    struct gl_shader **shader_list,
+                    unsigned num_shaders,
+                    struct gl_uniform_block **blocks_ret)
+{
+   /* This hash table will track all of the uniform blocks that have been
+    * encountered.  Since blocks with the same block-name must be the same,
+    * the hash is organized by block-name.
+    */
+   struct hash_table *block_hash =
+      _mesa_hash_table_create(mem_ctx, _mesa_key_hash_string,
+                              _mesa_key_string_equal);
+
+   if (block_hash == NULL) {
+      _mesa_error_no_memory(__func__);
+      linker_error(prog, "out of memory\n");
+      return 0;
+   }
+
+   /* Determine which uniform blocks are active.
+    */
+   link_uniform_block_active_visitor v(mem_ctx, block_hash, prog);
+   for (unsigned i = 0; i < num_shaders; i++) {
+      visit_list_elements(&v, shader_list[i]->ir);
+   }
+
+   /* Count the number of active uniform blocks.  Count the total number of
+    * active slots in those uniform blocks.
+    */
+   unsigned num_blocks = 0;
+   unsigned num_variables = 0;
+   count_block_size block_size;
+   struct hash_entry *entry;
+
+   hash_table_foreach (block_hash, entry) {
+      struct link_uniform_block_active *const b =
+         (struct link_uniform_block_active *) entry->data;
+
+      assert((b->array != NULL) == b->type->is_array());
+
+      if (b->array != NULL &&
+          (b->type->without_array()->interface_packing ==
+           GLSL_INTERFACE_PACKING_PACKED)) {
+         b->type = resize_block_array(b->type, b->array);
+         b->var->type = b->type;
+      }
+
+      block_size.num_active_uniforms = 0;
+      block_size.process(b->type->without_array(), "");
+
+      if (b->array != NULL) {
+         unsigned aoa_size = b->type->arrays_of_arrays_size();
+         num_blocks += aoa_size;
+         num_variables += aoa_size * block_size.num_active_uniforms;
+      } else {
+         num_blocks++;
+         num_variables += block_size.num_active_uniforms;
+      }
+
+   }
+
+   if (num_blocks == 0) {
+      assert(num_variables == 0);
+      _mesa_hash_table_destroy(block_hash, NULL);
+      return 0;
+   }
+
+   assert(num_variables != 0);
+
+   /* Allocate storage to hold all of the informatation related to uniform
+    * blocks that can be queried through the API.
+    */
+   gl_uniform_block *blocks =
+      ralloc_array(mem_ctx, gl_uniform_block, num_blocks);
+   gl_uniform_buffer_variable *variables =
+      ralloc_array(blocks, gl_uniform_buffer_variable, num_variables);
+
+   /* Add each variable from each uniform block to the API tracking
+    * structures.
+    */
+   unsigned i = 0;
+   ubo_visitor parcel(blocks, variables, num_variables);
+
+   STATIC_ASSERT(unsigned(GLSL_INTERFACE_PACKING_STD140)
+                 == unsigned(ubo_packing_std140));
+   STATIC_ASSERT(unsigned(GLSL_INTERFACE_PACKING_SHARED)
+                 == unsigned(ubo_packing_shared));
+   STATIC_ASSERT(unsigned(GLSL_INTERFACE_PACKING_PACKED)
+                 == unsigned(ubo_packing_packed));
+   STATIC_ASSERT(unsigned(GLSL_INTERFACE_PACKING_STD430)
+                 == unsigned(ubo_packing_std430));
+
+   hash_table_foreach (block_hash, entry) {
+      const struct link_uniform_block_active *const b =
+         (const struct link_uniform_block_active *) entry->data;
+      const glsl_type *block_type = b->type;
+
+      if (b->array != NULL) {
+         unsigned binding_offset = 0;
+         char *name = ralloc_strdup(NULL, block_type->without_array()->name);
+         size_t name_length = strlen(name);
+
+         assert(b->has_instance_name);
+         process_block_array(b->array, &name, name_length, blocks, &parcel,
+                             variables, b, &i, &binding_offset, ctx, prog);
+         ralloc_free(name);
+      } else {
+         blocks[i].Name = ralloc_strdup(blocks, block_type->name);
+         blocks[i].Uniforms = &variables[parcel.index];
+         blocks[i].Binding = (b->has_binding) ? b->binding : 0;
+         blocks[i].UniformBufferSize = 0;
+         blocks[i]._Packing =
+            gl_uniform_block_packing(block_type->interface_packing);
+
+         parcel.process(block_type,
+                        b->has_instance_name ? block_type->name : "");
+
+         blocks[i].UniformBufferSize = parcel.buffer_size;
+
+         /* Check SSBO size is lower than maximum supported size for SSBO */
+         if (b->is_shader_storage &&
+             parcel.buffer_size > ctx->Const.MaxShaderStorageBlockSize) {
+            linker_error(prog, "shader storage block `%s' has size %d, "
+                         "which is larger than than the maximum allowed (%d)",
+                         block_type->name,
+                         parcel.buffer_size,
+                         ctx->Const.MaxShaderStorageBlockSize);
+         }
+         blocks[i].NumUniforms =
+            (unsigned)(ptrdiff_t)(&variables[parcel.index] - blocks[i].Uniforms);
+
+         blocks[i].IsShaderStorage = b->is_shader_storage;
+
+         i++;
+      }
+   }
+
+   assert(parcel.index == num_variables);
+
+   _mesa_hash_table_destroy(block_hash, NULL);
+
+   *blocks_ret = blocks;
+   return num_blocks;
+}
+
 bool
 link_uniform_blocks_are_compatible(const gl_uniform_block *a,
                                   const gl_uniform_block *b)