[Mesa-dev] [PATCH v5 49/70] glsl: lower SSBO atomic intrinsics

[Iago Toral Quiroga]

The first argument to SSBO atomics is a reference to a SSBO buffer variable
so we want to compute its block index and offset and provide these values
to an internal version of the intrinsic that takes them instead of the
buffer variable reference.

v2:
- Support single components of integer vectors to be passed in as arguments.
- Get interface packing information from interface's type.
---
 src/glsl/lower_ubo_reference.cpp | 159 +++++++++++++++++++++++++++++++++++++++
 1 file changed, 159 insertions(+)

diff --git a/src/glsl/lower_ubo_reference.cpp b/src/glsl/lower_ubo_reference.cpp
index 368dbfa..2d77680 100644
--- a/src/glsl/lower_ubo_reference.cpp
+++ b/src/glsl/lower_ubo_reference.cpp
@@ -180,6 +180,10 @@ public:
    unsigned calculate_unsized_array_stride(ir_dereference *deref,
                                            unsigned packing);
 
+   ir_call *lower_ssbo_atomic_intrinsic(ir_call *ir);
+   ir_call *check_for_ssbo_atomic_intrinsic(ir_call *ir);
+   ir_visitor_status visit_enter(ir_call *ir);
+
    void *mem_ctx;
    struct gl_shader *shader;
    struct gl_uniform_buffer_variable *ubo_var;
@@ -242,7 +246,12 @@ interface_field_name(void *mem_ctx, char *base_name, ir_rvalue *d,
 
          break;
       }
+      case ir_type_swizzle: {
+         ir_swizzle *s = (ir_swizzle *) d;
 
+         d = s->val->as_dereference();
+         break;
+      }
       default:
          assert(!"Should not get here.");
          break;
@@ -427,6 +436,16 @@ lower_ubo_reference_visitor::setup_for_load_or_store(ir_variable *var,
          break;
       }
 
+      case ir_type_swizzle: {
+         ir_swizzle *deref_swizzle = (ir_swizzle *) deref;
+
+         assert(deref_swizzle->mask.num_components == 1);
+
+         *const_offset += deref_swizzle->mask.x * sizeof(int);
+         deref = deref_swizzle->val->as_dereference();
+         break;
+      }
+
       default:
          assert(!"not reached");
          deref = NULL;
@@ -1017,6 +1036,146 @@ lower_ubo_reference_visitor::visit_enter(ir_assignment *ir)
    return rvalue_visit(ir);
 }
 
+/* Lowers the intrinsic call to a new internal intrinsic that swaps the
+ * access to the buffer variable in the first parameter by an offset
+ * and block index. This involves creating the new internal intrinsic
+ * (i.e. the new function signature).
+ */
+ir_call *
+lower_ubo_reference_visitor::lower_ssbo_atomic_intrinsic(ir_call *ir)
+{
+   /* SSBO atomics usually have 2 parameters, the buffer variable and an
+    * integer argument. The exception is CompSwap, that has an additional
+    * integer parameter.
+    */
+   int param_count = ir->actual_parameters.length();
+   assert(param_count == 2 || param_count == 3);
+
+   /* First argument must be a scalar integer buffer variable */
+   exec_node *param = ir->actual_parameters.get_head();
+   ir_instruction *inst = (ir_instruction *) param;
+   assert(inst->ir_type == ir_type_dereference_variable ||
+          inst->ir_type == ir_type_dereference_array ||
+          inst->ir_type == ir_type_dereference_record ||
+          inst->ir_type == ir_type_swizzle);
+
+   ir_rvalue *deref = (ir_rvalue *) inst;
+   assert(deref->type->is_scalar() && deref->type->is_integer());
+
+   ir_variable *var = deref->variable_referenced();
+   assert(var);
+
+   /* Compute the offset to the start if the dereference and the
+    * block index
+    */
+   mem_ctx = ralloc_parent(shader->ir);
+
+   ir_rvalue *offset = NULL;
+   unsigned const_offset;
+   bool row_major;
+   int matrix_columns;
+   unsigned packing = var->get_interface_type()->interface_packing;
+
+   setup_for_load_or_store(var, deref,
+                           &offset, &const_offset,
+                           &row_major, &matrix_columns,
+                           packing);
+   assert(offset);
+   assert(!row_major);
+   assert(matrix_columns == 1);
+
+   ir_rvalue *deref_offset =
+      add(offset, new(mem_ctx) ir_constant(const_offset));
+   ir_rvalue *block_index = this->uniform_block->clone(mem_ctx, NULL);
+
+   /* Create the new internal function signature that will take a block
+    * index and offset instead of a buffer variable
+    */
+   exec_list sig_params;
+   ir_variable *sig_param = new(mem_ctx)
+      ir_variable(glsl_type::uint_type, "block_ref" , ir_var_function_in);
+   sig_params.push_tail(sig_param);
+
+   sig_param = new(mem_ctx)
+      ir_variable(glsl_type::uint_type, "offset" , ir_var_function_in);
+   sig_params.push_tail(sig_param);
+
+   const glsl_type *type = deref->type->base_type == GLSL_TYPE_INT ?
+      glsl_type::int_type : glsl_type::uint_type;
+   param = param->get_next();
+   sig_param = new(mem_ctx)
+         ir_variable(type, "data1", ir_var_function_in);
+   sig_params.push_tail(sig_param);
+
+   if (param_count == 3) {
+      param = param->get_next();
+      sig_param = new(mem_ctx)
+            ir_variable(type, "data2", ir_var_function_in);
+      sig_params.push_tail(sig_param);
+   }
+
+   ir_function_signature *sig =
+      new(mem_ctx) ir_function_signature(deref->type,
+                                         shader_storage_buffer_object);
+   assert(sig);
+   sig->replace_parameters(&sig_params);
+   sig->is_intrinsic = true;
+
+   char func_name[64];
+   sprintf(func_name, "%s_internal", ir->callee_name());
+   ir_function *f = new(mem_ctx) ir_function(func_name);
+   f->add_signature(sig);
+
+   /* Now, create the call to the internal intrinsic */
+   exec_list call_params;
+   call_params.push_tail(block_index);
+   call_params.push_tail(deref_offset);
+   param = ir->actual_parameters.get_head()->get_next();
+   ir_rvalue *param_as_rvalue = ((ir_instruction *) param)->as_rvalue();
+   call_params.push_tail(param_as_rvalue->clone(mem_ctx, NULL));
+   if (param_count == 3) {
+      param = param->get_next();
+      param_as_rvalue = ((ir_instruction *) param)->as_rvalue();
+      call_params.push_tail(param_as_rvalue->clone(mem_ctx, NULL));
+   }
+   ir_dereference_variable *return_deref =
+      ir->return_deref->clone(mem_ctx, NULL);
+   return new(mem_ctx) ir_call(sig, return_deref, &call_params);
+}
+
+ir_call *
+lower_ubo_reference_visitor::check_for_ssbo_atomic_intrinsic(ir_call *ir)
+{
+   const char *callee = ir->callee_name();
+   if (!strcmp("__intrinsic_ssbo_atomic_add", callee) ||
+       !strcmp("__intrinsic_ssbo_atomic_min", callee) ||
+       !strcmp("__intrinsic_ssbo_atomic_max", callee) ||
+       !strcmp("__intrinsic_ssbo_atomic_and", callee) ||
+       !strcmp("__intrinsic_ssbo_atomic_or", callee) ||
+       !strcmp("__intrinsic_ssbo_atomic_xor", callee) ||
+       !strcmp("__intrinsic_ssbo_atomic_exchange", callee) ||
+       !strcmp("__intrinsic_ssbo_atomic_comp_swap", callee)) {
+      return lower_ssbo_atomic_intrinsic(ir);
+   }
+
+   return ir;
+}
+
+
+ir_visitor_status
+lower_ubo_reference_visitor::visit_enter(ir_call *ir)
+{
+   ir_call *new_ir = check_for_ssbo_atomic_intrinsic(ir);
+   if (new_ir != ir) {
+      progress = true;
+      base_ir->replace_with(new_ir);
+      return visit_continue_with_parent;
+   }
+
+   return rvalue_visit(ir);
+}
+
+
 } /* unnamed namespace */
 
 void
-- 
1.9.1