[Mesa-dev] [PATCH 27/74] glsl: Lower shader storage buffer object writes to ir_ssbo_store

Jordan Justen jordan.l.justen at intel.com
Mon May 18 15:58:33 PDT 2015


On 2015-05-14 07:06:30, Iago Toral Quiroga wrote:
> Extend the existing lower_ubo_reference pass to also detect SSBO writes
> and lower them to ir_ssbo_store nodes.
> ---
>  src/glsl/lower_ubo_reference.cpp | 412 +++++++++++++++++++++++++++++++--------
>  1 file changed, 336 insertions(+), 76 deletions(-)
> 
> diff --git a/src/glsl/lower_ubo_reference.cpp b/src/glsl/lower_ubo_reference.cpp
> index fb8a890..8cb28f5 100644
> --- a/src/glsl/lower_ubo_reference.cpp
> +++ b/src/glsl/lower_ubo_reference.cpp
> @@ -139,12 +139,34 @@ public:
>     }
>  
>     void handle_rvalue(ir_rvalue **rvalue);
> +   ir_visitor_status visit_enter(ir_assignment *ir);
> +
> +   void setup_for_load_or_write(ir_variable *var,
> +                                ir_dereference *deref,
> +                                ir_rvalue **offset,
> +                                unsigned *const_offset,
> +                                bool *row_major,
> +                                int *matrix_columns,
> +                                bool *is_buffer);
> +
>     void emit_ubo_loads(ir_dereference *deref, ir_variable *base_offset,
>                         unsigned int deref_offset, bool row_major,
>                         int matrix_columns);
>     ir_expression *ubo_load(const struct glsl_type *type,
>                            ir_rvalue *offset);
>  
> +
> +   void check_for_ssbo_write(ir_assignment *ir);
> +   void write_to_memory(ir_dereference *deref,
> +                        ir_variable *var,
> +                        ir_variable *write_var,
> +                        unsigned write_mask);
> +   void emit_ssbo_writes(ir_dereference *deref, ir_variable *base_offset,
> +                         unsigned int deref_offset, bool row_major,
> +                         int matrix_columns, unsigned write_mask);
> +   ir_ssbo_store *ssbo_write(ir_rvalue *deref, ir_rvalue *offset,
> +                             unsigned write_mask);
> +
>     void *mem_ctx;
>     struct gl_shader *shader;
>     struct gl_uniform_buffer_variable *ubo_var;
> @@ -219,26 +241,21 @@ interface_field_name(void *mem_ctx, char *base_name, ir_dereference *d,
>  }
>  
>  void
> -lower_ubo_reference_visitor::handle_rvalue(ir_rvalue **rvalue)
> +lower_ubo_reference_visitor::setup_for_load_or_write(ir_variable *var,
> +                                                     ir_dereference *deref,
> +                                                     ir_rvalue **offset,
> +                                                     unsigned *const_offset,
> +                                                     bool *row_major,
> +                                                     int *matrix_columns,
> +                                                     bool *is_buffer)
>  {
> -   if (!*rvalue)
> -      return;
> -
> -   ir_dereference *deref = (*rvalue)->as_dereference();
> -   if (!deref)
> -      return;
> -
> -   ir_variable *var = deref->variable_referenced();
> -   if (!var || !var->is_in_uniform_block())
> -      return;
> -
> -   mem_ctx = ralloc_parent(*rvalue);
> -
> +   /* Fix out the name of the interface block*/
>     ir_rvalue *nonconst_block_index;
>     const char *const field_name =
>        interface_field_name(mem_ctx, (char *) var->get_interface_type()->name,
>                             deref, &nonconst_block_index);
>  
> +   /* Locate the ubo block by interface name */
>     this->uniform_block = NULL;
>     for (unsigned i = 0; i < shader->NumUniformBlocks; i++) {
>        if (strcmp(field_name, shader->UniformBlocks[i].Name) == 0) {
> @@ -253,10 +270,7 @@ lower_ubo_reference_visitor::handle_rvalue(ir_rvalue **rvalue)
>              this->uniform_block = index;
>           }
>  
> -         if (shader->UniformBlocks[i].IsBuffer)
> -            this->opcode = ir_binop_ssbo_load;
> -         else
> -            this->opcode = ir_binop_ubo_load;
> +         *is_buffer = shader->UniformBlocks[i].IsBuffer;
>  
>           struct gl_uniform_block *block = &shader->UniformBlocks[i];
>  
> @@ -269,10 +283,10 @@ lower_ubo_reference_visitor::handle_rvalue(ir_rvalue **rvalue)
>  
>     assert(this->uniform_block);
>  
> -   ir_rvalue *offset = new(mem_ctx) ir_constant(0u);
> -   unsigned const_offset = 0;
> -   bool row_major = is_dereferenced_thing_row_major(deref);
> -   int matrix_columns = 1;
> +   *offset = new(mem_ctx) ir_constant(0u);
> +   *const_offset = 0;
> +   *row_major = is_dereferenced_thing_row_major(deref);
> +   *matrix_columns = 1;
>  
>     /* Calculate the offset to the start of the region of the UBO
>      * dereferenced by *rvalue.  This may be a variable offset if an
> @@ -281,76 +295,76 @@ lower_ubo_reference_visitor::handle_rvalue(ir_rvalue **rvalue)
>     while (deref) {
>        switch (deref->ir_type) {
>        case ir_type_dereference_variable: {
> -        const_offset += ubo_var->Offset;
> -        deref = NULL;
> -        break;
> +         *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;
> +         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;
>              if (deref_array->array->type->is_double())
>                 array_stride *= 2;
> -            matrix_columns = deref_array->array->type->matrix_columns;
> +            *matrix_columns = deref_array->array->type->matrix_columns;
>           } else if (deref_array->type->is_interface()) {
>              /* We're processing an array dereference of an interface instance
> -            * array.  The thing being dereferenced *must* be a variable
> -            * dereference because intefaces cannot be embedded an other
> -            * types.  In terms of calculating the offsets for the lowering
> -            * pass, we don't care about the array index.  All elements of an
> -            * interface instance array will have the same offsets relative to
> -            * the base of the block that backs them.
> +             * array. The thing being dereferenced *must* be a variable
> +             * dereference because interfaces cannot be embedded in other
> +             * types. In terms of calculating the offsets for the lowering
> +             * pass, we don't care about the array index. All elements of an
> +             * interface instance array will have the same offsets relative to
> +             * the base of the block that backs them.
>               */
>              assert(deref_array->array->as_dereference_variable());
>              deref = deref_array->array->as_dereference();
>              break;
> -        } else {
> +         } else {
>              /* Whether or not the field is row-major (because it might be a
> -             * bvec2 or something) does not affect the array itself.  We need
> +             * bvec2 or something) does not affect the array itself. We need
>               * to know whether an array element in its entirety is row-major.
>               */
>              const bool array_row_major =
>                 is_dereferenced_thing_row_major(deref_array);
>  
> -           array_stride = deref_array->type->std140_size(array_row_major);
> -           array_stride = glsl_align(array_stride, 16);
> -        }
> +            array_stride = deref_array->type->std140_size(array_row_major);
> +            array_stride = glsl_align(array_stride, 16);
> +         }
>  
>           ir_rvalue *array_index = deref_array->array_index;
>           if (array_index->type->base_type == GLSL_TYPE_INT)
>              array_index = i2u(array_index);
>  
> -        ir_constant *const_index =
> +         ir_constant *const_index =
>              array_index->constant_expression_value(NULL);
> -        if (const_index) {
> -           const_offset += array_stride * const_index->value.u[0];
> -        } else {
> -           offset = add(offset,
> -                        mul(array_index,
> -                            new(mem_ctx) ir_constant(array_stride)));
> -        }
> -        deref = deref_array->array->as_dereference();
> -        break;
> +         if (const_index) {
> +            *const_offset += array_stride * const_index->value.u[0];
> +         } else {
> +            *offset = add(*offset,
> +                          mul(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;
> +         ir_dereference_record *deref_record = (ir_dereference_record *) deref;
> +         const glsl_type *struct_type = deref_record->record->type;
> +         unsigned intra_struct_offset = 0;
>  
> -        for (unsigned int i = 0; i < struct_type->length; i++) {
> -           const glsl_type *type = struct_type->fields.structure[i].type;
> +         for (unsigned int i = 0; i < struct_type->length; i++) {
> +            const glsl_type *type = struct_type->fields.structure[i].type;
>  
> -            ir_dereference_record *field_deref =
> -               new(mem_ctx) ir_dereference_record(deref_record->record,
> -                                                  struct_type->fields.structure[i].name);
> +            ir_dereference_record *field_deref = new(mem_ctx)
> +               ir_dereference_record(deref_record->record,
> +                                     struct_type->fields.structure[i].name);
>              const bool field_row_major =
>                 is_dereferenced_thing_row_major(field_deref);
>  
> @@ -358,11 +372,12 @@ lower_ubo_reference_visitor::handle_rvalue(ir_rvalue **rvalue)
>  
>              unsigned field_align = type->std140_base_alignment(field_row_major);
>  
> -           intra_struct_offset = glsl_align(intra_struct_offset, field_align);
> +            intra_struct_offset = glsl_align(intra_struct_offset, field_align);
> +
> +            if (strcmp(struct_type->fields.structure[i].name,
> +                       deref_record->field) == 0)
> +               break;
>  
> -           if (strcmp(struct_type->fields.structure[i].name,
> -                      deref_record->field) == 0)
> -              break;
>              intra_struct_offset += type->std140_size(field_row_major);
>  
>              /* If the field just examined was itself a structure, apply rule
> @@ -377,19 +392,56 @@ lower_ubo_reference_visitor::handle_rvalue(ir_rvalue **rvalue)
>                                                  field_align);
>  
>              }
> -        }
> -
> -        const_offset += intra_struct_offset;
> +         }
>  
> -        deref = deref_record->record->as_dereference();
> -        break;
> +         *const_offset += intra_struct_offset;
> +         deref = deref_record->record->as_dereference();
> +         break;
>        }
> +
>        default:
> -        assert(!"not reached");
> -        deref = NULL;
> -        break;
> +         assert(!"not reached");
> +         deref = NULL;
> +         break;
>        }
>     }
> +}
> +
> +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->is_in_uniform_block())
> +      return;
> +
> +   mem_ctx = ralloc_parent(*rvalue);
> +
> +   ir_rvalue *offset = NULL;
> +   unsigned const_offset;
> +   bool row_major;
> +   int matrix_columns;
> +   bool is_buffer;
> +
> +   /* Compute the offset to the start if the dereference as well as other
> +    * information we need to configure the write
> +    */
> +   setup_for_load_or_write(var, deref,
> +                           &offset, &const_offset,
> +                           &row_major, &matrix_columns,
> +                           &is_buffer);
> +   assert(offset);
> +
> +   if (is_buffer)
> +      this->opcode = ir_binop_ssbo_load;
> +   else
> +      this->opcode = ir_binop_ubo_load;
>  
>     /* Now that we've calculated the offset to the start of the
>      * dereference, walk over the type and emit loads into a temporary.
> @@ -549,6 +601,214 @@ lower_ubo_reference_visitor::emit_ubo_loads(ir_dereference *deref,
>     }
>  }
>  
> +ir_ssbo_store *
> +lower_ubo_reference_visitor::ssbo_write(ir_rvalue *deref,
> +                                        ir_rvalue *offset,
> +                                        unsigned write_mask)
> +{
> +   ir_rvalue *block_ref = this->uniform_block->clone(mem_ctx, NULL);
> +   ir_rvalue *val_ref = deref->clone(mem_ctx, NULL);
> +   return new(mem_ctx) ir_ssbo_store(block_ref, offset, val_ref, write_mask);
> +}
> +
> +static inline int
> +writemask_for_size(unsigned n)
> +{
> +   return ((1 << n) - 1);
> +}
> +
> +void
> +lower_ubo_reference_visitor::emit_ssbo_writes(ir_dereference *deref,
> +                                              ir_variable *base_offset,
> +                                              unsigned int deref_offset,
> +                                              bool row_major,
> +                                              int matrix_columns,
> +                                              unsigned write_mask)
> +{

This code seems to be mostly a duplicate of the load case. Wouldn't
this leave us with two mostly similar complex code paths?

Maybe the low level emit could switch based on the opcode?

-Jordan

> +   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 =
> +            glsl_align(field_offset,
> +                       field->type->std140_base_alignment(row_major));
> +
> +         emit_ssbo_writes(field_deref, base_offset,
> +                          deref_offset + field_offset,
> +                          row_major, 1,
> +                          writemask_for_size(field_deref->type->vector_elements));
> +
> +             field_offset += field->type->std140_size(row_major);
> +      }
> +      return;
> +   }
> +
> +   if (deref->type->is_array()) {
> +      unsigned array_stride =
> +         glsl_align(deref->type->fields.array->std140_size(row_major), 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_ssbo_writes(element_deref, base_offset,
> +                          deref_offset + i * array_stride,
> +                          row_major, 1,
> +                          writemask_for_size(element_deref->type->vector_elements));
> +      }
> +      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);
> +
> +         if (row_major) {
> +            /* For a row-major matrix, the next column starts at the next
> +             * element.
> +             */
> +            int size_mul = deref->type->is_double() ? 8 : 4;
> +            emit_ssbo_writes(col_deref, base_offset,
> +                             deref_offset + i * size_mul,
> +                             row_major, deref->type->matrix_columns,
> +                             writemask_for_size(col_deref->type->vector_elements));
> +         } else {
> +            /* std140 always rounds the stride of arrays (and matrices) to a
> +             * vec4, so matrices are always 16 between columns/rows. With
> +             * doubles, they will be 32 apart when there are more than 2 rows.
> +             */
> +            int size_mul = (deref->type->is_double() &&
> +                            deref->type->vector_elements > 2) ? 32 : 16;
> +            emit_ssbo_writes(col_deref, base_offset,
> +                             deref_offset + i * size_mul,
> +                             row_major, deref->type->matrix_columns,
> +                             writemask_for_size(col_deref->type->vector_elements));
> +         }
> +      }
> +      return;
> +   }
> +
> +   assert(deref->type->is_scalar() || deref->type->is_vector());
> +
> +   if (!row_major) {
> +      ir_rvalue *offset =
> +         add(base_offset, new(mem_ctx) ir_constant(deref_offset));
> +      base_ir->insert_after(ssbo_write(deref, offset, write_mask));
> +   } else {
> +      unsigned N = deref->type->is_double() ? 8 : 4;
> +
> +      /* 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 ||
> +             deref->type->base_type == GLSL_TYPE_DOUBLE);
> +      /* 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. However a double matrix may either be 16
> +       * or 32 depending on the number of columns.
> +       */
> +      assert(matrix_columns <= 4);
> +      unsigned matrix_stride = glsl_align(matrix_columns * N, 16);
> +
> +      for (unsigned i = 0; i < deref->type->vector_elements; i++) {
> +         ir_rvalue *chan_offset =
> +            add(base_offset,
> +                new(mem_ctx) ir_constant(deref_offset + i * matrix_stride));
> +
> +         base_ir->insert_after(ssbo_write(swizzle(deref, i, 1), chan_offset, 1));
> +      }
> +   }
> +}
> +
> +void
> +lower_ubo_reference_visitor::write_to_memory(ir_dereference *deref,
> +                                             ir_variable *var,
> +                                             ir_variable *write_var,
> +                                             unsigned write_mask)
> +{
> +   ir_rvalue *offset = NULL;
> +   unsigned const_offset;
> +   bool row_major;
> +   int matrix_columns;
> +   bool is_buffer;
> +
> +   /* Compute the offset to the start if the dereference as well as other
> +    * information we need to configure the write
> +    */
> +   setup_for_load_or_write(var, deref,
> +                           &offset, &const_offset,
> +                           &row_major, &matrix_columns,
> +                           &is_buffer);
> +   assert(offset);
> +   assert(is_buffer);
> +
> +   /* Now emit writes from the temporary to memory */
> +   ir_variable *write_offset =
> +      new(mem_ctx) ir_variable(glsl_type::uint_type,
> +                               "ssbo_write_temp_offset",
> +                               ir_var_temporary);
> +
> +   base_ir->insert_before(write_offset);
> +   base_ir->insert_before(assign(write_offset, offset));
> +
> +   deref = new(mem_ctx) ir_dereference_variable(write_var);
> +   emit_ssbo_writes(deref, write_offset, const_offset,
> +                    row_major, matrix_columns, write_mask);
> +}
> +
> +void
> +lower_ubo_reference_visitor::check_for_ssbo_write(ir_assignment *ir)
> +{
> +   if (!ir || !ir->lhs)
> +      return;
> +
> +   ir_rvalue *rvalue = ir->lhs->as_rvalue();
> +   if (!rvalue)
> +      return;
> +
> +   ir_dereference *deref = ir->lhs->as_dereference();
> +   if (!deref)
> +      return;
> +
> +   ir_variable *var = ir->lhs->variable_referenced();
> +   if (!var || !var->is_in_uniform_block())
> +      return;
> +
> +   /* We have a write to a buffer variable, so declare a temporary and rewrite
> +    * the assignment so that the temporary is the LHS.
> +    */
> +   mem_ctx = ralloc_parent(rvalue);
> +
> +   const glsl_type *type = rvalue->type;
> +   ir_variable *write_var = new(mem_ctx) ir_variable(type,
> +                                                     "ssbo_write_temp",
> +                                                     ir_var_temporary);
> +   base_ir->insert_before(write_var);
> +   ir->lhs = new(mem_ctx) ir_dereference_variable(write_var);
> +
> +   /* Now we have to write the value assigned to the temporary back to memory */
> +   write_to_memory(deref, var, write_var, ir->write_mask);
> +   progress = true;
> +}
> +
> +
> +ir_visitor_status
> +lower_ubo_reference_visitor::visit_enter(ir_assignment *ir)
> +{
> +   check_for_ssbo_write(ir);
> +   return rvalue_visit(ir);
> +}
> +
>  } /* unnamed namespace */
>  
>  void
> -- 
> 1.9.1
> 
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev


More information about the mesa-dev mailing list