glsl: Add a lowering pass to turn complicated UBO references to vector loads.
authorEric Anholt <eric@anholt.net>
Wed, 11 Jul 2012 15:26:31 +0000 (08:26 -0700)
committerEric Anholt <eric@anholt.net>
Tue, 7 Aug 2012 20:54:47 +0000 (13:54 -0700)
v2: Reduce the impenetrable code in emit_ubo_loads() by 23 lines by keeping
    the ir_variable as the variable part of the offset from handle_rvalue(),
    and track the constant offsets from that with a plain old integer value,
    avoiding a bunch of temporary variables in the array and struct handling.
    Also, fix file description doxygen.
v3: Fix a row vs col typo, and fix spelling in a comment.

Reviewed-by: Eric Anholt <eric@anholt.net>
src/glsl/Makefile.sources
src/glsl/ir_optimization.h
src/glsl/lower_ubo_reference.cpp [new file with mode: 0644]

index f2743f750670ce80aa5411ae9f14fbd6250b1d07..765f06a27cbb8b23e222e82f7e33f9a139a8120f 100644 (file)
@@ -66,6 +66,7 @@ LIBGLSL_CXX_FILES = \
        $(GLSL_SRCDIR)/lower_vec_index_to_swizzle.cpp \
        $(GLSL_SRCDIR)/lower_vector.cpp \
        $(GLSL_SRCDIR)/lower_output_reads.cpp \
+       $(GLSL_SRCDIR)/lower_ubo_reference.cpp \
        $(GLSL_SRCDIR)/opt_algebraic.cpp \
        $(GLSL_SRCDIR)/opt_array_splitting.cpp \
        $(GLSL_SRCDIR)/opt_constant_folding.cpp \
index c435d7717261d010181555902660c94ebff9abd4..2220d511e96cb5123f89ae5ad13eaa329b40ab4c 100644 (file)
@@ -74,6 +74,7 @@ bool lower_variable_index_to_cond_assign(exec_list *instructions,
 bool lower_quadop_vector(exec_list *instructions, bool dont_lower_swz);
 bool lower_clip_distance(exec_list *instructions);
 void lower_output_reads(exec_list *instructions);
+void lower_ubo_reference(struct gl_shader *shader, exec_list *instructions);
 bool optimize_redundant_jumps(exec_list *instructions);
 bool optimize_split_arrays(exec_list *instructions, bool linked);
 
diff --git a/src/glsl/lower_ubo_reference.cpp b/src/glsl/lower_ubo_reference.cpp
new file mode 100644 (file)
index 0000000..e8d2c47
--- /dev/null
@@ -0,0 +1,313 @@
+/*
+ * Copyright © 2012 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ */
+
+/**
+ * \file lower_ubo_reference.cpp
+ *
+ * IR lower pass to replace dereferences of variables in a uniform
+ * buffer object with usage of ir_binop_ubo_load expressions, each of
+ * which can read data up to the size of a vec4.
+ *
+ * This relieves drivers of the responsibility to deal with tricky UBO
+ * layout issues like std140 structures and row_major matrices on
+ * their own.
+ */
+
+#include "ir.h"
+#include "ir_builder.h"
+#include "ir_rvalue_visitor.h"
+#include "main/macros.h"
+
+using namespace ir_builder;
+
+namespace {
+class lower_ubo_reference_visitor : public ir_rvalue_enter_visitor {
+public:
+   lower_ubo_reference_visitor(struct gl_shader *shader)
+   : shader(shader)
+   {
+   }
+
+   void handle_rvalue(ir_rvalue **rvalue);
+   void emit_ubo_loads(ir_dereference *deref, ir_variable *base_offset,
+                      unsigned int deref_offset);
+   ir_expression *ubo_load(const struct glsl_type *type,
+                          ir_rvalue *offset);
+
+   void *mem_ctx;
+   struct gl_shader *shader;
+   struct gl_uniform_buffer_variable *ubo_var;
+   unsigned uniform_block;
+   bool progress;
+};
+
+static inline unsigned int
+align(unsigned int a, unsigned int align)
+{
+   return (a + align - 1) / align * align;
+}
+
+void
+lower_ubo_reference_visitor::handle_rvalue(ir_rvalue **rvalue)
+{
+   if (!*rvalue)
+      return;
+
+   ir_dereference *deref = (*rvalue)->as_dereference();
+   if (!deref)
+      return;
+
+   ir_variable *var = deref->variable_referenced();
+   if (!var || var->uniform_block == -1)
+      return;
+
+   mem_ctx = ralloc_parent(*rvalue);
+   uniform_block = var->uniform_block;
+   struct gl_uniform_block *block = &shader->UniformBlocks[uniform_block];
+   this->ubo_var = &block->Uniforms[var->location];
+   ir_rvalue *offset = new(mem_ctx) ir_constant(0u);
+   unsigned const_offset = 0;
+   bool row_major = ubo_var->RowMajor;
+
+   /* Calculate the offset to the start of the region of the UBO
+    * dereferenced by *rvalue.  This may be a variable offset if an
+    * array dereference has a variable index.
+    */
+   while (deref) {
+      switch (deref->ir_type) {
+      case ir_type_dereference_variable: {
+        const_offset += ubo_var->Offset;
+        deref = NULL;
+        break;
+      }
+
+      case ir_type_dereference_array: {
+        ir_dereference_array *deref_array = (ir_dereference_array *)deref;
+        unsigned array_stride;
+        if (deref_array->array->type->is_matrix() && row_major) {
+           /* When loading a vector out of a row major matrix, the
+            * step between the columns (vectors) is the size of a
+            * float, while the step between the rows (elements of a
+            * vector) is handled below in emit_ubo_loads.
+            */
+           array_stride = 4;
+        } else {
+           array_stride = deref_array->type->std140_size(row_major);
+           array_stride = align(array_stride, 16);
+        }
+
+        ir_constant *const_index = deref_array->array_index->as_constant();
+        if (const_index) {
+           const_offset += array_stride * const_index->value.i[0];
+        } else {
+           offset = add(offset,
+                        mul(deref_array->array_index,
+                            new(mem_ctx) ir_constant(array_stride)));
+        }
+        deref = deref_array->array->as_dereference();
+        break;
+      }
+
+      case ir_type_dereference_record: {
+        ir_dereference_record *deref_record = (ir_dereference_record *)deref;
+        const glsl_type *struct_type = deref_record->record->type;
+        unsigned intra_struct_offset = 0;
+
+        unsigned max_field_align = 16;
+        for (unsigned int i = 0; i < struct_type->length; i++) {
+           const glsl_type *type = struct_type->fields.structure[i].type;
+           unsigned field_align = type->std140_base_alignment(row_major);
+           max_field_align = MAX2(field_align, max_field_align);
+           intra_struct_offset = align(intra_struct_offset, field_align);
+
+           if (strcmp(struct_type->fields.structure[i].name,
+                      deref_record->field) == 0)
+              break;
+           intra_struct_offset += type->std140_size(row_major);
+        }
+
+        const_offset = align(const_offset, max_field_align);
+        const_offset += intra_struct_offset;
+
+        deref = deref_record->record->as_dereference();
+        break;
+      }
+      default:
+        assert(!"not reached");
+        deref = NULL;
+        break;
+      }
+   }
+
+   /* Now that we've calculated the offset to the start of the
+    * dereference, walk over the type and emit loads into a temporary.
+    */
+   const glsl_type *type = (*rvalue)->type;
+   ir_variable *load_var = new(mem_ctx) ir_variable(type,
+                                                   "ubo_load_temp",
+                                                   ir_var_temporary);
+   base_ir->insert_before(load_var);
+
+   ir_variable *load_offset = new(mem_ctx) ir_variable(glsl_type::uint_type,
+                                                      "ubo_load_temp_offset",
+                                                      ir_var_temporary);
+   base_ir->insert_before(load_offset);
+   base_ir->insert_before(assign(load_offset, offset));
+
+   deref = new(mem_ctx) ir_dereference_variable(load_var);
+   emit_ubo_loads(deref, load_offset, const_offset);
+   *rvalue = deref;
+
+   progress = true;
+}
+
+ir_expression *
+lower_ubo_reference_visitor::ubo_load(const glsl_type *type,
+                                     ir_rvalue *offset)
+{
+   return new(mem_ctx)
+      ir_expression(ir_binop_ubo_load,
+                   type,
+                   new(mem_ctx) ir_constant(this->uniform_block),
+                   offset);
+
+}
+
+/**
+ * Takes LHS and emits a series of assignments into its components
+ * from the UBO variable at variable_offset + deref_offset.
+ *
+ * Recursively calls itself to break the deref down to the point that
+ * the ir_binop_ubo_load expressions generated are contiguous scalars
+ * or vectors.
+ */
+void
+lower_ubo_reference_visitor::emit_ubo_loads(ir_dereference *deref,
+                                           ir_variable *base_offset,
+                                           unsigned int deref_offset)
+{
+   if (deref->type->is_record()) {
+      unsigned int field_offset = 0;
+
+      for (unsigned i = 0; i < deref->type->length; i++) {
+        const struct glsl_struct_field *field =
+           &deref->type->fields.structure[i];
+        ir_dereference *field_deref =
+           new(mem_ctx) ir_dereference_record(deref->clone(mem_ctx, NULL),
+                                              field->name);
+
+        field_offset =
+           align(field_offset,
+                 field->type->std140_base_alignment(ubo_var->RowMajor));
+
+        emit_ubo_loads(field_deref, base_offset, deref_offset + field_offset);
+
+        field_offset += field->type->std140_size(ubo_var->RowMajor);
+      }
+      return;
+   }
+
+   if (deref->type->is_array()) {
+      unsigned array_stride =
+        align(deref->type->fields.array->std140_size(ubo_var->RowMajor), 16);
+
+      for (unsigned i = 0; i < deref->type->length; i++) {
+        ir_constant *element = new(mem_ctx) ir_constant(i);
+        ir_dereference *element_deref =
+           new(mem_ctx) ir_dereference_array(deref->clone(mem_ctx, NULL),
+                                             element);
+        emit_ubo_loads(element_deref, base_offset,
+                       deref_offset + i * array_stride);
+      }
+      return;
+   }
+
+   if (deref->type->is_matrix()) {
+      for (unsigned i = 0; i < deref->type->matrix_columns; i++) {
+        ir_constant *col = new(mem_ctx) ir_constant(i);
+        ir_dereference *col_deref =
+           new(mem_ctx) ir_dereference_array(deref->clone(mem_ctx, NULL),
+                                             col);
+
+        /* std140 always rounds the stride of arrays (and matrices)
+         * to a vec4, so matrices are always 16 between columns/rows.
+         */
+        emit_ubo_loads(col_deref, base_offset, deref_offset + i * 16);
+      }
+      return;
+   }
+
+   assert(deref->type->is_scalar() ||
+         deref->type->is_vector());
+
+   if (!ubo_var->RowMajor) {
+      ir_rvalue *offset = add(base_offset,
+                             new(mem_ctx) ir_constant(deref_offset));
+      base_ir->insert_before(assign(deref->clone(mem_ctx, NULL),
+                                   ubo_load(deref->type, offset)));
+   } else {
+      /* We're dereffing a column out of a row-major matrix, so we
+       * gather the vector from each stored row.
+      */
+      assert(deref->type->base_type == GLSL_TYPE_FLOAT);
+      /* Matrices, row_major or not, are stored as if they were
+       * arrays of vectors of the appropriate size in std140.
+       * Arrays have their strides rounded up to a vec4, so the
+       * matrix stride is always 16.
+       */
+      unsigned matrix_stride = 16;
+
+      for (unsigned i = 0; i < deref->type->vector_elements; i++) {
+        ir_rvalue *chan = new(mem_ctx) ir_constant((int)i);
+        ir_dereference *deref_chan =
+           new(mem_ctx) ir_dereference_array(deref->clone(mem_ctx, NULL),
+                                             chan);
+
+        ir_rvalue *chan_offset =
+           add(base_offset,
+               new(mem_ctx) ir_constant(deref_offset + i * matrix_stride));
+
+        base_ir->insert_before(assign(deref_chan,
+                                      ubo_load(glsl_type::float_type,
+                                               chan_offset)));
+      }
+   }
+}
+
+} /* unnamed namespace */
+
+void
+lower_ubo_reference(struct gl_shader *shader, exec_list *instructions)
+{
+   lower_ubo_reference_visitor v(shader);
+
+   /* Loop over the instructions lowering references, because we take
+    * a deref of a UBO array using a UBO dereference as the index will
+    * produce a collection of instructions all of which have cloned
+    * UBO dereferences for that array index.
+    */
+   do {
+      v.progress = false;
+      visit_list_elements(&v, instructions);
+   } while (v.progress);
+}