[Mesa-dev] [PATCH] glsl: Add a lowering pass to turn complicated UBO references to vector loads.
Kenneth Graunke
kenneth at whitecape.org
Tue Aug 7 11:45:12 PDT 2012
On 08/06/2012 07:00 PM, Eric Anholt wrote:
> 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.
> ---
> src/glsl/Makefile.sources | 1 +
> src/glsl/ir_optimization.h | 1 +
> src/glsl/lower_ubo_reference.cpp | 313 ++++++++++++++++++++++++++++++++++++++
> 3 files changed, 315 insertions(+)
> create mode 100644 src/glsl/lower_ubo_reference.cpp
>
> diff --git a/src/glsl/Makefile.sources b/src/glsl/Makefile.sources
> index f2743f7..765f06a 100644
> --- a/src/glsl/Makefile.sources
> +++ b/src/glsl/Makefile.sources
> @@ -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 \
> diff --git a/src/glsl/ir_optimization.h b/src/glsl/ir_optimization.h
> index c435d77..2220d51 100644
> --- a/src/glsl/ir_optimization.h
> +++ b/src/glsl/ir_optimization.h
> @@ -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
> index 0000000..f930da5
> --- /dev/null
> +++ b/src/glsl/lower_ubo_reference.cpp
> @@ -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 derefernece has a variable index.
Typo: derefernece -> dereference
> + */
> + 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 *row = new(mem_ctx) ir_constant(i);
> + ir_dereference *row_deref =
> + new(mem_ctx) ir_dereference_array(deref->clone(mem_ctx, NULL),
> + row);
Pretty sure these aren't rows. They're cols. Well. I suppose they
could be either. Argh. I was going to suggest renaming them to "cols",
but...debatable whether that's better.
The offset stuff is much nicer in v2, thanks!
As is, this patch is:
Reviewed-by: Kenneth Graunke <kenneth at whitecape.org>
> +
> + /* std140 always rounds the stride of arrays (and matrices)
> + * to a vec4, so matrices are always 16 between columns/rows.
> + */
> + emit_ubo_loads(row_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);
> +}
>
More information about the mesa-dev
mailing list