i965: remove now unused GLSL IR optimisations
These are no longer used since the previous commit. Acked-by: Elie Tournier <elie.tournier@collabora.com> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This commit is contained in:
parent
ad55b1a770
commit
794ae44095
|
@ -28,8 +28,6 @@ i965_FILES = \
|
|||
brw_ff_gs.c \
|
||||
brw_ff_gs_emit.c \
|
||||
brw_ff_gs.h \
|
||||
brw_fs_channel_expressions.cpp \
|
||||
brw_fs_vector_splitting.cpp \
|
||||
brw_formatquery.c \
|
||||
brw_gs.c \
|
||||
brw_gs.h \
|
||||
|
|
|
@ -1,483 +0,0 @@
|
|||
/*
|
||||
* Copyright © 2010 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 brw_wm_channel_expressions.cpp
|
||||
*
|
||||
* Breaks vector operations down into operations on each component.
|
||||
*
|
||||
* The 965 fragment shader receives 8 or 16 pixels at a time, so each
|
||||
* channel of a vector is laid out as 1 or 2 8-float registers. Each
|
||||
* ALU operation operates on one of those channel registers. As a
|
||||
* result, there is no value to the 965 fragment shader in tracking
|
||||
* "vector" expressions in the sense of GLSL fragment shaders, when
|
||||
* doing a channel at a time may help in constant folding, algebraic
|
||||
* simplification, and reducing the liveness of channel registers.
|
||||
*
|
||||
* The exception to the desire to break everything down to floats is
|
||||
* texturing. The texture sampler returns a writemasked masked
|
||||
* 4/8-register sequence containing the texture values. We don't want
|
||||
* to dispatch to the sampler separately for each channel we need, so
|
||||
* we do retain the vector types in that case.
|
||||
*/
|
||||
|
||||
#include "brw_program.h"
|
||||
#include "compiler/glsl/ir.h"
|
||||
#include "compiler/glsl/ir_expression_flattening.h"
|
||||
#include "compiler/glsl_types.h"
|
||||
|
||||
class ir_channel_expressions_visitor : public ir_hierarchical_visitor {
|
||||
public:
|
||||
ir_channel_expressions_visitor()
|
||||
{
|
||||
this->progress = false;
|
||||
this->mem_ctx = NULL;
|
||||
}
|
||||
|
||||
ir_visitor_status visit_leave(ir_assignment *);
|
||||
|
||||
ir_rvalue *get_element(ir_variable *var, unsigned int element);
|
||||
void assign(ir_assignment *ir, int elem, ir_rvalue *val);
|
||||
|
||||
bool progress;
|
||||
void *mem_ctx;
|
||||
};
|
||||
|
||||
static bool
|
||||
channel_expressions_predicate(ir_instruction *ir)
|
||||
{
|
||||
ir_expression *expr = ir->as_expression();
|
||||
unsigned int i;
|
||||
|
||||
if (!expr)
|
||||
return false;
|
||||
|
||||
switch (expr->operation) {
|
||||
case ir_unop_pack_half_2x16:
|
||||
case ir_unop_pack_snorm_2x16:
|
||||
case ir_unop_pack_snorm_4x8:
|
||||
case ir_unop_pack_unorm_2x16:
|
||||
case ir_unop_pack_unorm_4x8:
|
||||
return false;
|
||||
|
||||
/* these opcodes need to act on the whole vector,
|
||||
* just like texturing.
|
||||
*/
|
||||
case ir_unop_interpolate_at_centroid:
|
||||
case ir_binop_interpolate_at_offset:
|
||||
case ir_binop_interpolate_at_sample:
|
||||
case ir_unop_pack_double_2x32:
|
||||
case ir_unop_pack_int_2x32:
|
||||
case ir_unop_pack_uint_2x32:
|
||||
return false;
|
||||
default:
|
||||
break;
|
||||
}
|
||||
|
||||
for (i = 0; i < expr->get_num_operands(); i++) {
|
||||
if (expr->operands[i]->type->is_vector())
|
||||
return true;
|
||||
}
|
||||
|
||||
return false;
|
||||
}
|
||||
|
||||
bool
|
||||
brw_do_channel_expressions(exec_list *instructions)
|
||||
{
|
||||
ir_channel_expressions_visitor v;
|
||||
|
||||
/* Pull out any matrix expression to a separate assignment to a
|
||||
* temp. This will make our handling of the breakdown to
|
||||
* operations on the matrix's vector components much easier.
|
||||
*/
|
||||
do_expression_flattening(instructions, channel_expressions_predicate);
|
||||
|
||||
visit_list_elements(&v, instructions);
|
||||
|
||||
return v.progress;
|
||||
}
|
||||
|
||||
ir_rvalue *
|
||||
ir_channel_expressions_visitor::get_element(ir_variable *var, unsigned int elem)
|
||||
{
|
||||
ir_dereference *deref;
|
||||
|
||||
if (var->type->is_scalar())
|
||||
return new(mem_ctx) ir_dereference_variable(var);
|
||||
|
||||
assert(elem < var->type->components());
|
||||
deref = new(mem_ctx) ir_dereference_variable(var);
|
||||
return new(mem_ctx) ir_swizzle(deref, elem, 0, 0, 0, 1);
|
||||
}
|
||||
|
||||
void
|
||||
ir_channel_expressions_visitor::assign(ir_assignment *ir, int elem, ir_rvalue *val)
|
||||
{
|
||||
ir_dereference *lhs = ir->lhs->clone(mem_ctx, NULL);
|
||||
ir_assignment *assign;
|
||||
|
||||
/* This assign-of-expression should have been generated by the
|
||||
* expression flattening visitor (since we never short circit to
|
||||
* not flatten, even for plain assignments of variables), so the
|
||||
* writemask is always full.
|
||||
*/
|
||||
assert(ir->write_mask == (1 << ir->lhs->type->components()) - 1);
|
||||
|
||||
assign = new(mem_ctx) ir_assignment(lhs, val, NULL, (1 << elem));
|
||||
ir->insert_before(assign);
|
||||
}
|
||||
|
||||
ir_visitor_status
|
||||
ir_channel_expressions_visitor::visit_leave(ir_assignment *ir)
|
||||
{
|
||||
ir_expression *expr = ir->rhs->as_expression();
|
||||
bool found_vector = false;
|
||||
unsigned int i, vector_elements = 1;
|
||||
ir_variable *op_var[4];
|
||||
|
||||
if (!expr)
|
||||
return visit_continue;
|
||||
|
||||
if (!this->mem_ctx)
|
||||
this->mem_ctx = ralloc_parent(ir);
|
||||
|
||||
for (i = 0; i < expr->get_num_operands(); i++) {
|
||||
if (expr->operands[i]->type->is_vector()) {
|
||||
found_vector = true;
|
||||
vector_elements = expr->operands[i]->type->vector_elements;
|
||||
break;
|
||||
}
|
||||
}
|
||||
if (!found_vector)
|
||||
return visit_continue;
|
||||
|
||||
switch (expr->operation) {
|
||||
case ir_unop_pack_half_2x16:
|
||||
case ir_unop_pack_snorm_2x16:
|
||||
case ir_unop_pack_snorm_4x8:
|
||||
case ir_unop_pack_unorm_2x16:
|
||||
case ir_unop_pack_unorm_4x8:
|
||||
case ir_unop_interpolate_at_centroid:
|
||||
case ir_binop_interpolate_at_offset:
|
||||
case ir_binop_interpolate_at_sample:
|
||||
/* We scalarize these in NIR, so no need to do it here */
|
||||
case ir_unop_pack_double_2x32:
|
||||
case ir_unop_pack_int_2x32:
|
||||
case ir_unop_pack_uint_2x32:
|
||||
return visit_continue;
|
||||
|
||||
default:
|
||||
break;
|
||||
}
|
||||
|
||||
/* Store the expression operands in temps so we can use them
|
||||
* multiple times.
|
||||
*/
|
||||
for (i = 0; i < expr->get_num_operands(); i++) {
|
||||
ir_assignment *assign;
|
||||
ir_dereference *deref;
|
||||
|
||||
assert(!expr->operands[i]->type->is_matrix());
|
||||
|
||||
op_var[i] = new(mem_ctx) ir_variable(expr->operands[i]->type,
|
||||
"channel_expressions",
|
||||
ir_var_temporary);
|
||||
ir->insert_before(op_var[i]);
|
||||
|
||||
deref = new(mem_ctx) ir_dereference_variable(op_var[i]);
|
||||
assign = new(mem_ctx) ir_assignment(deref,
|
||||
expr->operands[i],
|
||||
NULL);
|
||||
ir->insert_before(assign);
|
||||
}
|
||||
|
||||
const glsl_type *element_type = glsl_type::get_instance(ir->lhs->type->base_type,
|
||||
1, 1);
|
||||
|
||||
/* OK, time to break down this vector operation. */
|
||||
switch (expr->operation) {
|
||||
case ir_unop_bit_not:
|
||||
case ir_unop_logic_not:
|
||||
case ir_unop_neg:
|
||||
case ir_unop_abs:
|
||||
case ir_unop_sign:
|
||||
case ir_unop_rcp:
|
||||
case ir_unop_rsq:
|
||||
case ir_unop_sqrt:
|
||||
case ir_unop_exp:
|
||||
case ir_unop_log:
|
||||
case ir_unop_exp2:
|
||||
case ir_unop_log2:
|
||||
case ir_unop_bitcast_i2f:
|
||||
case ir_unop_bitcast_f2i:
|
||||
case ir_unop_bitcast_f2u:
|
||||
case ir_unop_bitcast_u2f:
|
||||
case ir_unop_bitcast_u642d:
|
||||
case ir_unop_bitcast_i642d:
|
||||
case ir_unop_bitcast_d2u64:
|
||||
case ir_unop_bitcast_d2i64:
|
||||
case ir_unop_i2u:
|
||||
case ir_unop_u2i:
|
||||
case ir_unop_f2i:
|
||||
case ir_unop_f2u:
|
||||
case ir_unop_i2f:
|
||||
case ir_unop_f2b:
|
||||
case ir_unop_b2f:
|
||||
case ir_unop_i2b:
|
||||
case ir_unop_b2i:
|
||||
case ir_unop_u2f:
|
||||
case ir_unop_d2f:
|
||||
case ir_unop_f2d:
|
||||
case ir_unop_d2i:
|
||||
case ir_unop_i2d:
|
||||
case ir_unop_d2u:
|
||||
case ir_unop_u2d:
|
||||
case ir_unop_d2b:
|
||||
case ir_unop_i642i:
|
||||
case ir_unop_u642i:
|
||||
case ir_unop_i642u:
|
||||
case ir_unop_u642u:
|
||||
case ir_unop_i642b:
|
||||
case ir_unop_i642f:
|
||||
case ir_unop_u642f:
|
||||
case ir_unop_i642d:
|
||||
case ir_unop_u642d:
|
||||
case ir_unop_i2i64:
|
||||
case ir_unop_u2i64:
|
||||
case ir_unop_b2i64:
|
||||
case ir_unop_f2i64:
|
||||
case ir_unop_d2i64:
|
||||
case ir_unop_i2u64:
|
||||
case ir_unop_u2u64:
|
||||
case ir_unop_f2u64:
|
||||
case ir_unop_d2u64:
|
||||
case ir_unop_u642i64:
|
||||
case ir_unop_i642u64:
|
||||
case ir_unop_trunc:
|
||||
case ir_unop_ceil:
|
||||
case ir_unop_floor:
|
||||
case ir_unop_fract:
|
||||
case ir_unop_round_even:
|
||||
case ir_unop_sin:
|
||||
case ir_unop_cos:
|
||||
case ir_unop_dFdx:
|
||||
case ir_unop_dFdx_coarse:
|
||||
case ir_unop_dFdx_fine:
|
||||
case ir_unop_dFdy:
|
||||
case ir_unop_dFdy_coarse:
|
||||
case ir_unop_dFdy_fine:
|
||||
case ir_unop_bitfield_reverse:
|
||||
case ir_unop_bit_count:
|
||||
case ir_unop_find_msb:
|
||||
case ir_unop_find_lsb:
|
||||
case ir_unop_saturate:
|
||||
case ir_unop_subroutine_to_int:
|
||||
for (i = 0; i < vector_elements; i++) {
|
||||
ir_rvalue *op0 = get_element(op_var[0], i);
|
||||
|
||||
assign(ir, i, new(mem_ctx) ir_expression(expr->operation,
|
||||
element_type,
|
||||
op0,
|
||||
NULL));
|
||||
}
|
||||
break;
|
||||
|
||||
case ir_binop_add:
|
||||
case ir_binop_sub:
|
||||
case ir_binop_mul:
|
||||
case ir_binop_imul_high:
|
||||
case ir_binop_div:
|
||||
case ir_binop_carry:
|
||||
case ir_binop_borrow:
|
||||
case ir_binop_mod:
|
||||
case ir_binop_min:
|
||||
case ir_binop_max:
|
||||
case ir_binop_pow:
|
||||
case ir_binop_lshift:
|
||||
case ir_binop_rshift:
|
||||
case ir_binop_bit_and:
|
||||
case ir_binop_bit_xor:
|
||||
case ir_binop_bit_or:
|
||||
case ir_binop_logic_and:
|
||||
case ir_binop_logic_xor:
|
||||
case ir_binop_logic_or:
|
||||
case ir_binop_less:
|
||||
case ir_binop_greater:
|
||||
case ir_binop_lequal:
|
||||
case ir_binop_gequal:
|
||||
case ir_binop_equal:
|
||||
case ir_binop_nequal:
|
||||
case ir_binop_ldexp:
|
||||
for (i = 0; i < vector_elements; i++) {
|
||||
ir_rvalue *op0 = get_element(op_var[0], i);
|
||||
ir_rvalue *op1 = get_element(op_var[1], i);
|
||||
|
||||
assign(ir, i, new(mem_ctx) ir_expression(expr->operation,
|
||||
element_type,
|
||||
op0,
|
||||
op1));
|
||||
}
|
||||
break;
|
||||
|
||||
case ir_binop_dot: {
|
||||
ir_expression *last = NULL;
|
||||
for (i = 0; i < vector_elements; i++) {
|
||||
ir_rvalue *op0 = get_element(op_var[0], i);
|
||||
ir_rvalue *op1 = get_element(op_var[1], i);
|
||||
ir_expression *temp;
|
||||
|
||||
temp = new(mem_ctx) ir_expression(ir_binop_mul,
|
||||
element_type,
|
||||
op0,
|
||||
op1);
|
||||
if (last) {
|
||||
last = new(mem_ctx) ir_expression(ir_binop_add,
|
||||
element_type,
|
||||
temp,
|
||||
last);
|
||||
} else {
|
||||
last = temp;
|
||||
}
|
||||
}
|
||||
assign(ir, 0, last);
|
||||
break;
|
||||
}
|
||||
|
||||
case ir_binop_all_equal:
|
||||
case ir_binop_any_nequal: {
|
||||
ir_expression *last = NULL;
|
||||
for (i = 0; i < vector_elements; i++) {
|
||||
ir_rvalue *op0 = get_element(op_var[0], i);
|
||||
ir_rvalue *op1 = get_element(op_var[1], i);
|
||||
ir_expression *temp;
|
||||
ir_expression_operation join;
|
||||
|
||||
if (expr->operation == ir_binop_all_equal)
|
||||
join = ir_binop_logic_and;
|
||||
else
|
||||
join = ir_binop_logic_or;
|
||||
|
||||
temp = new(mem_ctx) ir_expression(expr->operation,
|
||||
element_type,
|
||||
op0,
|
||||
op1);
|
||||
if (last) {
|
||||
last = new(mem_ctx) ir_expression(join,
|
||||
element_type,
|
||||
temp,
|
||||
last);
|
||||
} else {
|
||||
last = temp;
|
||||
}
|
||||
}
|
||||
assign(ir, 0, last);
|
||||
break;
|
||||
}
|
||||
case ir_unop_noise:
|
||||
unreachable("noise should have been broken down to function call");
|
||||
|
||||
case ir_binop_ubo_load:
|
||||
case ir_unop_get_buffer_size:
|
||||
unreachable("not yet supported");
|
||||
|
||||
case ir_triop_fma:
|
||||
case ir_triop_lrp:
|
||||
case ir_triop_csel:
|
||||
case ir_triop_bitfield_extract:
|
||||
for (i = 0; i < vector_elements; i++) {
|
||||
ir_rvalue *op0 = get_element(op_var[0], i);
|
||||
ir_rvalue *op1 = get_element(op_var[1], i);
|
||||
ir_rvalue *op2 = get_element(op_var[2], i);
|
||||
|
||||
assign(ir, i, new(mem_ctx) ir_expression(expr->operation,
|
||||
element_type,
|
||||
op0,
|
||||
op1,
|
||||
op2));
|
||||
}
|
||||
break;
|
||||
|
||||
case ir_quadop_bitfield_insert:
|
||||
for (i = 0; i < vector_elements; i++) {
|
||||
ir_rvalue *op0 = get_element(op_var[0], i);
|
||||
ir_rvalue *op1 = get_element(op_var[1], i);
|
||||
ir_rvalue *op2 = get_element(op_var[2], i);
|
||||
ir_rvalue *op3 = get_element(op_var[3], i);
|
||||
|
||||
assign(ir, i, new(mem_ctx) ir_expression(expr->operation,
|
||||
element_type,
|
||||
op0,
|
||||
op1,
|
||||
op2,
|
||||
op3));
|
||||
}
|
||||
break;
|
||||
|
||||
case ir_unop_pack_snorm_2x16:
|
||||
case ir_unop_pack_snorm_4x8:
|
||||
case ir_unop_pack_unorm_2x16:
|
||||
case ir_unop_pack_unorm_4x8:
|
||||
case ir_unop_pack_half_2x16:
|
||||
case ir_unop_unpack_snorm_2x16:
|
||||
case ir_unop_unpack_snorm_4x8:
|
||||
case ir_unop_unpack_unorm_2x16:
|
||||
case ir_unop_unpack_unorm_4x8:
|
||||
case ir_unop_unpack_half_2x16:
|
||||
case ir_binop_vector_extract:
|
||||
case ir_triop_vector_insert:
|
||||
case ir_quadop_vector:
|
||||
case ir_unop_ssbo_unsized_array_length:
|
||||
unreachable("should have been lowered");
|
||||
|
||||
case ir_unop_interpolate_at_centroid:
|
||||
case ir_binop_interpolate_at_offset:
|
||||
case ir_binop_interpolate_at_sample:
|
||||
case ir_unop_unpack_double_2x32:
|
||||
unreachable("not reached: expression operates on scalars only");
|
||||
|
||||
case ir_unop_pack_double_2x32:
|
||||
case ir_unop_pack_int_2x32:
|
||||
case ir_unop_pack_uint_2x32:
|
||||
unreachable("not reached: to be lowered in NIR, should've been skipped");
|
||||
|
||||
case ir_unop_frexp_sig:
|
||||
case ir_unop_frexp_exp:
|
||||
unreachable("should have been lowered by lower_instructions");
|
||||
|
||||
case ir_unop_vote_any:
|
||||
case ir_unop_vote_all:
|
||||
case ir_unop_vote_eq:
|
||||
case ir_unop_unpack_int_2x32:
|
||||
case ir_unop_unpack_uint_2x32:
|
||||
case ir_unop_ballot:
|
||||
case ir_unop_read_first_invocation:
|
||||
case ir_binop_read_invocation:
|
||||
unreachable("unsupported");
|
||||
}
|
||||
|
||||
ir->remove();
|
||||
this->progress = true;
|
||||
|
||||
return visit_continue;
|
||||
}
|
|
@ -1,400 +0,0 @@
|
|||
/*
|
||||
* Copyright © 2010 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 brw_wm_vector_splitting.cpp
|
||||
*
|
||||
* If a vector is only ever referenced by its components, then
|
||||
* split those components out to individual variables so they can be
|
||||
* handled normally by other optimization passes.
|
||||
*
|
||||
* This skips vectors in uniforms and varyings, which need to be
|
||||
* accessible as vectors for their access by the GL. Also, vector
|
||||
* results of non-variable-derefs in assignments aren't handled
|
||||
* because to do so we would have to store the vector result to a
|
||||
* temporary in order to unload each channel, and to do so would just
|
||||
* loop us back to where we started. For the 965, this is exactly the
|
||||
* behavior we want for the results of texture lookups, but probably not for
|
||||
*/
|
||||
|
||||
#include "brw_program.h"
|
||||
#include "main/imports.h"
|
||||
#include "compiler/glsl/ir.h"
|
||||
#include "compiler/glsl/ir_rvalue_visitor.h"
|
||||
#include "compiler/glsl_types.h"
|
||||
#include "util/hash_table.h"
|
||||
|
||||
static bool debug = false;
|
||||
|
||||
class variable_entry : public exec_node
|
||||
{
|
||||
public:
|
||||
variable_entry(ir_variable *var)
|
||||
{
|
||||
this->var = var;
|
||||
this->whole_vector_access = 0;
|
||||
this->mem_ctx = NULL;
|
||||
}
|
||||
|
||||
ir_variable *var; /* The key: the variable's pointer. */
|
||||
|
||||
/** Number of times the variable is referenced, including assignments. */
|
||||
unsigned whole_vector_access;
|
||||
|
||||
ir_variable *components[4];
|
||||
|
||||
/** ralloc_parent(this->var) -- the shader's ralloc context. */
|
||||
void *mem_ctx;
|
||||
};
|
||||
|
||||
class ir_vector_reference_visitor : public ir_hierarchical_visitor {
|
||||
public:
|
||||
ir_vector_reference_visitor(void)
|
||||
{
|
||||
this->mem_ctx = ralloc_context(NULL);
|
||||
this->ht = _mesa_hash_table_create(mem_ctx, _mesa_hash_pointer,
|
||||
_mesa_key_pointer_equal);
|
||||
}
|
||||
|
||||
~ir_vector_reference_visitor(void)
|
||||
{
|
||||
ralloc_free(mem_ctx);
|
||||
}
|
||||
|
||||
virtual ir_visitor_status visit(ir_variable *);
|
||||
virtual ir_visitor_status visit(ir_dereference_variable *);
|
||||
virtual ir_visitor_status visit_enter(ir_swizzle *);
|
||||
virtual ir_visitor_status visit_enter(ir_assignment *);
|
||||
virtual ir_visitor_status visit_enter(ir_function_signature *);
|
||||
|
||||
variable_entry *get_variable_entry(ir_variable *var);
|
||||
|
||||
/* List of variable_entry */
|
||||
struct hash_table *ht;
|
||||
|
||||
void *mem_ctx;
|
||||
};
|
||||
|
||||
variable_entry *
|
||||
ir_vector_reference_visitor::get_variable_entry(ir_variable *var)
|
||||
{
|
||||
assert(var);
|
||||
|
||||
if (!var->type->is_vector())
|
||||
return NULL;
|
||||
|
||||
switch (var->data.mode) {
|
||||
case ir_var_uniform:
|
||||
case ir_var_shader_storage:
|
||||
case ir_var_shader_shared:
|
||||
case ir_var_shader_in:
|
||||
case ir_var_shader_out:
|
||||
case ir_var_system_value:
|
||||
case ir_var_function_in:
|
||||
case ir_var_function_out:
|
||||
case ir_var_function_inout:
|
||||
/* Can't split varyings or uniforms. Function in/outs won't get split
|
||||
* either.
|
||||
*/
|
||||
return NULL;
|
||||
case ir_var_auto:
|
||||
case ir_var_temporary:
|
||||
break;
|
||||
}
|
||||
|
||||
struct hash_entry *hte = _mesa_hash_table_search(ht, var);
|
||||
if (hte)
|
||||
return (struct variable_entry *) hte->data;
|
||||
|
||||
variable_entry *entry = new(mem_ctx) variable_entry(var);
|
||||
_mesa_hash_table_insert(ht, var, entry);
|
||||
return entry;
|
||||
}
|
||||
|
||||
|
||||
ir_visitor_status
|
||||
ir_vector_reference_visitor::visit(ir_variable *ir)
|
||||
{
|
||||
/* Make sure splitting looks at splitting this variable */
|
||||
(void)this->get_variable_entry(ir);
|
||||
|
||||
return visit_continue;
|
||||
}
|
||||
|
||||
ir_visitor_status
|
||||
ir_vector_reference_visitor::visit(ir_dereference_variable *ir)
|
||||
{
|
||||
ir_variable *const var = ir->var;
|
||||
variable_entry *entry = this->get_variable_entry(var);
|
||||
|
||||
if (entry)
|
||||
entry->whole_vector_access++;
|
||||
|
||||
return visit_continue;
|
||||
}
|
||||
|
||||
ir_visitor_status
|
||||
ir_vector_reference_visitor::visit_enter(ir_swizzle *ir)
|
||||
{
|
||||
/* Don't descend into a vector ir_dereference_variable below. */
|
||||
if (ir->val->as_dereference_variable() && ir->type->is_scalar())
|
||||
return visit_continue_with_parent;
|
||||
|
||||
return visit_continue;
|
||||
}
|
||||
|
||||
ir_visitor_status
|
||||
ir_vector_reference_visitor::visit_enter(ir_assignment *ir)
|
||||
{
|
||||
if (ir->lhs->as_dereference_variable() &&
|
||||
ir->rhs->as_dereference_variable() &&
|
||||
!ir->condition) {
|
||||
/* We'll split copies of a vector to copies of channels, so don't
|
||||
* descend to the ir_dereference_variables.
|
||||
*/
|
||||
return visit_continue_with_parent;
|
||||
}
|
||||
if (ir->lhs->as_dereference_variable() &&
|
||||
_mesa_is_pow_two(ir->write_mask) &&
|
||||
!ir->condition) {
|
||||
/* If we're writing just a channel, then channel-splitting the LHS is OK.
|
||||
*/
|
||||
ir->rhs->accept(this);
|
||||
return visit_continue_with_parent;
|
||||
}
|
||||
return visit_continue;
|
||||
}
|
||||
|
||||
ir_visitor_status
|
||||
ir_vector_reference_visitor::visit_enter(ir_function_signature *ir)
|
||||
{
|
||||
/* We don't want to descend into the function parameters and
|
||||
* split them, so just accept the body here.
|
||||
*/
|
||||
visit_list_elements(this, &ir->body);
|
||||
return visit_continue_with_parent;
|
||||
}
|
||||
|
||||
class ir_vector_splitting_visitor : public ir_rvalue_visitor {
|
||||
public:
|
||||
ir_vector_splitting_visitor(struct hash_table *vars)
|
||||
{
|
||||
this->ht = vars;
|
||||
}
|
||||
|
||||
virtual ir_visitor_status visit_leave(ir_assignment *);
|
||||
|
||||
void handle_rvalue(ir_rvalue **rvalue);
|
||||
variable_entry *get_splitting_entry(ir_variable *var);
|
||||
|
||||
struct hash_table *ht;
|
||||
};
|
||||
|
||||
variable_entry *
|
||||
ir_vector_splitting_visitor::get_splitting_entry(ir_variable *var)
|
||||
{
|
||||
assert(var);
|
||||
|
||||
if (!var->type->is_vector())
|
||||
return NULL;
|
||||
|
||||
struct hash_entry *hte = _mesa_hash_table_search(ht, var);
|
||||
return hte ? (struct variable_entry *) hte->data : NULL;
|
||||
}
|
||||
|
||||
void
|
||||
ir_vector_splitting_visitor::handle_rvalue(ir_rvalue **rvalue)
|
||||
{
|
||||
if (!*rvalue)
|
||||
return;
|
||||
|
||||
ir_swizzle *swiz = (*rvalue)->as_swizzle();
|
||||
if (!swiz || !swiz->type->is_scalar())
|
||||
return;
|
||||
|
||||
ir_dereference_variable *deref_var = swiz->val->as_dereference_variable();
|
||||
if (!deref_var)
|
||||
return;
|
||||
|
||||
variable_entry *entry = get_splitting_entry(deref_var->var);
|
||||
if (!entry)
|
||||
return;
|
||||
|
||||
ir_variable *var = entry->components[swiz->mask.x];
|
||||
*rvalue = new(entry->mem_ctx) ir_dereference_variable(var);
|
||||
}
|
||||
|
||||
ir_visitor_status
|
||||
ir_vector_splitting_visitor::visit_leave(ir_assignment *ir)
|
||||
{
|
||||
ir_dereference_variable *lhs_deref = ir->lhs->as_dereference_variable();
|
||||
ir_dereference_variable *rhs_deref = ir->rhs->as_dereference_variable();
|
||||
variable_entry *lhs = lhs_deref ? get_splitting_entry(lhs_deref->var) : NULL;
|
||||
variable_entry *rhs = rhs_deref ? get_splitting_entry(rhs_deref->var) : NULL;
|
||||
|
||||
if (lhs_deref && rhs_deref && (lhs || rhs) && !ir->condition) {
|
||||
unsigned int rhs_chan = 0;
|
||||
|
||||
/* Straight assignment of vector variables. */
|
||||
for (unsigned int i = 0; i < ir->lhs->type->vector_elements; i++) {
|
||||
ir_dereference *new_lhs;
|
||||
ir_rvalue *new_rhs;
|
||||
void *mem_ctx = lhs ? lhs->mem_ctx : rhs->mem_ctx;
|
||||
unsigned int writemask;
|
||||
|
||||
if (!(ir->write_mask & (1 << i)))
|
||||
continue;
|
||||
|
||||
if (lhs) {
|
||||
new_lhs = new(mem_ctx) ir_dereference_variable(lhs->components[i]);
|
||||
writemask = 1;
|
||||
} else {
|
||||
new_lhs = ir->lhs->clone(mem_ctx, NULL);
|
||||
writemask = 1 << i;
|
||||
}
|
||||
|
||||
if (rhs) {
|
||||
new_rhs =
|
||||
new(mem_ctx) ir_dereference_variable(rhs->components[rhs_chan]);
|
||||
} else {
|
||||
new_rhs = new(mem_ctx) ir_swizzle(ir->rhs->clone(mem_ctx, NULL),
|
||||
rhs_chan, 0, 0, 0, 1);
|
||||
}
|
||||
|
||||
ir->insert_before(new(mem_ctx) ir_assignment(new_lhs,
|
||||
new_rhs,
|
||||
NULL, writemask));
|
||||
|
||||
rhs_chan++;
|
||||
}
|
||||
ir->remove();
|
||||
} else if (lhs) {
|
||||
void *mem_ctx = lhs->mem_ctx;
|
||||
int elem = -1;
|
||||
|
||||
switch (ir->write_mask) {
|
||||
case (1 << 0):
|
||||
elem = 0;
|
||||
break;
|
||||
case (1 << 1):
|
||||
elem = 1;
|
||||
break;
|
||||
case (1 << 2):
|
||||
elem = 2;
|
||||
break;
|
||||
case (1 << 3):
|
||||
elem = 3;
|
||||
break;
|
||||
default:
|
||||
ir->fprint(stderr);
|
||||
unreachable("not reached: non-channelwise dereference of LHS.");
|
||||
}
|
||||
|
||||
ir->lhs = new(mem_ctx) ir_dereference_variable(lhs->components[elem]);
|
||||
ir->write_mask = (1 << 0);
|
||||
|
||||
handle_rvalue(&ir->rhs);
|
||||
} else {
|
||||
handle_rvalue(&ir->rhs);
|
||||
}
|
||||
|
||||
handle_rvalue(&ir->condition);
|
||||
|
||||
return visit_continue;
|
||||
}
|
||||
|
||||
bool
|
||||
brw_do_vector_splitting(exec_list *instructions)
|
||||
{
|
||||
struct hash_entry *hte;
|
||||
|
||||
ir_vector_reference_visitor refs;
|
||||
|
||||
visit_list_elements(&refs, instructions);
|
||||
|
||||
/* Trim out variables we can't split. */
|
||||
hash_table_foreach(refs.ht, hte) {
|
||||
struct variable_entry *entry = (struct variable_entry *) hte->data;
|
||||
if (debug) {
|
||||
fprintf(stderr, "vector %s@%p: whole_access %d\n",
|
||||
entry->var->name, (void *) entry->var,
|
||||
entry->whole_vector_access);
|
||||
}
|
||||
|
||||
if (entry->whole_vector_access) {
|
||||
_mesa_hash_table_remove(refs.ht, hte);
|
||||
}
|
||||
}
|
||||
|
||||
if (refs.ht->entries == 0)
|
||||
return false;
|
||||
|
||||
void *mem_ctx = ralloc_context(NULL);
|
||||
|
||||
/* Replace the decls of the vectors to be split with their split
|
||||
* components.
|
||||
*/
|
||||
hash_table_foreach(refs.ht, hte) {
|
||||
struct variable_entry *entry = (struct variable_entry *) hte->data;
|
||||
const struct glsl_type *type;
|
||||
type = glsl_type::get_instance(entry->var->type->base_type, 1, 1);
|
||||
|
||||
entry->mem_ctx = ralloc_parent(entry->var);
|
||||
|
||||
for (unsigned int i = 0; i < entry->var->type->vector_elements; i++) {
|
||||
char *const name = ir_variable::temporaries_allocate_names
|
||||
? ralloc_asprintf(mem_ctx, "%s_%c",
|
||||
entry->var->name,
|
||||
"xyzw"[i])
|
||||
: NULL;
|
||||
|
||||
entry->components[i] = new(entry->mem_ctx) ir_variable(type, name,
|
||||
ir_var_temporary);
|
||||
|
||||
ralloc_free(name);
|
||||
|
||||
if (entry->var->constant_initializer) {
|
||||
ir_constant_data data = {0};
|
||||
assert(entry->var->data.has_initializer);
|
||||
if (entry->var->type->is_double()) {
|
||||
data.d[0] = entry->var->constant_initializer->value.d[i];
|
||||
} else {
|
||||
data.u[0] = entry->var->constant_initializer->value.u[i];
|
||||
}
|
||||
entry->components[i]->data.has_initializer = true;
|
||||
entry->components[i]->constant_initializer = new(entry->components[i]) ir_constant(type, &data);
|
||||
}
|
||||
|
||||
entry->var->insert_before(entry->components[i]);
|
||||
}
|
||||
|
||||
entry->var->remove();
|
||||
}
|
||||
|
||||
ir_vector_splitting_visitor split(refs.ht);
|
||||
visit_list_elements(&split, instructions);
|
||||
|
||||
ralloc_free(mem_ctx);
|
||||
|
||||
return true;
|
||||
}
|
|
@ -32,9 +32,6 @@ extern "C" {
|
|||
|
||||
struct brw_context;
|
||||
|
||||
bool brw_do_channel_expressions(struct exec_list *instructions);
|
||||
bool brw_do_vector_splitting(struct exec_list *instructions);
|
||||
|
||||
struct nir_shader *brw_create_nir(struct brw_context *brw,
|
||||
const struct gl_shader_program *shader_prog,
|
||||
struct gl_program *prog,
|
||||
|
|
Loading…
Reference in New Issue