aboutsummaryrefslogtreecommitdiff
path: root/mesalib/src/glsl/lower_instructions.cpp
diff options
context:
space:
mode:
authormarha <marha@users.sourceforge.net>2010-12-28 16:10:20 +0000
committermarha <marha@users.sourceforge.net>2010-12-28 16:10:20 +0000
commit807c6931fe683fd844ceec1b023232181e6aae03 (patch)
tree1a131ed95fe2200d0ad33da8f7755a7ed2364adc /mesalib/src/glsl/lower_instructions.cpp
parent973099dda7e49e5abe29819a7124b3b1e7bd8b92 (diff)
downloadvcxsrv-807c6931fe683fd844ceec1b023232181e6aae03.tar.gz
vcxsrv-807c6931fe683fd844ceec1b023232181e6aae03.tar.bz2
vcxsrv-807c6931fe683fd844ceec1b023232181e6aae03.zip
xserver and mesa git update 28-12-2010
Diffstat (limited to 'mesalib/src/glsl/lower_instructions.cpp')
-rw-r--r--mesalib/src/glsl/lower_instructions.cpp288
1 files changed, 288 insertions, 0 deletions
diff --git a/mesalib/src/glsl/lower_instructions.cpp b/mesalib/src/glsl/lower_instructions.cpp
new file mode 100644
index 000000000..6e44d1319
--- /dev/null
+++ b/mesalib/src/glsl/lower_instructions.cpp
@@ -0,0 +1,288 @@
+/*
+ * 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 lower_instructions.cpp
+ *
+ * Many GPUs lack native instructions for certain expression operations, and
+ * must replace them with some other expression tree. This pass lowers some
+ * of the most common cases, allowing the lowering code to be implemented once
+ * rather than in each driver backend.
+ *
+ * Currently supported transformations:
+ * - SUB_TO_ADD_NEG
+ * - DIV_TO_MUL_RCP
+ * - EXP_TO_EXP2
+ * - POW_TO_EXP2
+ * - LOG_TO_LOG2
+ * - MOD_TO_FRACT
+ *
+ * SUB_TO_ADD_NEG:
+ * ---------------
+ * Breaks an ir_binop_sub expression down to add(op0, neg(op1))
+ *
+ * This simplifies expression reassociation, and for many backends
+ * there is no subtract operation separate from adding the negation.
+ * For backends with native subtract operations, they will probably
+ * want to recognize add(op0, neg(op1)) or the other way around to
+ * produce a subtract anyway.
+ *
+ * DIV_TO_MUL_RCP:
+ * ---------------
+ * Breaks an ir_unop_div expression down to op0 * (rcp(op1)).
+ *
+ * Many GPUs don't have a divide instruction (945 and 965 included),
+ * but they do have an RCP instruction to compute an approximate
+ * reciprocal. By breaking the operation down, constant reciprocals
+ * can get constant folded.
+ *
+ * EXP_TO_EXP2 and LOG_TO_LOG2:
+ * ----------------------------
+ * Many GPUs don't have a base e log or exponent instruction, but they
+ * do have base 2 versions, so this pass converts exp and log to exp2
+ * and log2 operations.
+ *
+ * POW_TO_EXP2:
+ * -----------
+ * Many older GPUs don't have an x**y instruction. For these GPUs, convert
+ * x**y to 2**(y * log2(x)).
+ *
+ * MOD_TO_FRACT:
+ * -------------
+ * Breaks an ir_unop_mod expression down to (op1 * fract(op0 / op1))
+ *
+ * Many GPUs don't have a MOD instruction (945 and 965 included), and
+ * if we have to break it down like this anyway, it gives an
+ * opportunity to do things like constant fold the (1.0 / op1) easily.
+ */
+
+#include "main/core.h" /* for M_LOG2E */
+#include "glsl_types.h"
+#include "ir.h"
+#include "ir_optimization.h"
+
+class lower_instructions_visitor : public ir_hierarchical_visitor {
+public:
+ lower_instructions_visitor(unsigned lower)
+ : progress(false), lower(lower) { }
+
+ ir_visitor_status visit_leave(ir_expression *);
+
+ bool progress;
+
+private:
+ unsigned lower; /** Bitfield of which operations to lower */
+
+ void sub_to_add_neg(ir_expression *);
+ void div_to_mul_rcp(ir_expression *);
+ void mod_to_fract(ir_expression *);
+ void exp_to_exp2(ir_expression *);
+ void pow_to_exp2(ir_expression *);
+ void log_to_log2(ir_expression *);
+};
+
+/**
+ * Determine if a particular type of lowering should occur
+ */
+#define lowering(x) (this->lower & x)
+
+bool
+lower_instructions(exec_list *instructions, unsigned what_to_lower)
+{
+ lower_instructions_visitor v(what_to_lower);
+
+ visit_list_elements(&v, instructions);
+ return v.progress;
+}
+
+void
+lower_instructions_visitor::sub_to_add_neg(ir_expression *ir)
+{
+ ir->operation = ir_binop_add;
+ ir->operands[1] = new(ir) ir_expression(ir_unop_neg, ir->operands[1]->type,
+ ir->operands[1], NULL);
+ this->progress = true;
+}
+
+void
+lower_instructions_visitor::div_to_mul_rcp(ir_expression *ir)
+{
+ if (!ir->operands[1]->type->is_integer()) {
+ /* New expression for the 1.0 / op1 */
+ ir_rvalue *expr;
+ expr = new(ir) ir_expression(ir_unop_rcp,
+ ir->operands[1]->type,
+ ir->operands[1],
+ NULL);
+
+ /* op0 / op1 -> op0 * (1.0 / op1) */
+ ir->operation = ir_binop_mul;
+ ir->operands[1] = expr;
+ } else {
+ /* Be careful with integer division -- we need to do it as a
+ * float and re-truncate, since rcp(n > 1) of an integer would
+ * just be 0.
+ */
+ ir_rvalue *op0, *op1;
+ const struct glsl_type *vec_type;
+
+ vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
+ ir->operands[1]->type->vector_elements,
+ ir->operands[1]->type->matrix_columns);
+
+ if (ir->operands[1]->type->base_type == GLSL_TYPE_INT)
+ op1 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[1], NULL);
+ else
+ op1 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[1], NULL);
+
+ op1 = new(ir) ir_expression(ir_unop_rcp, op1->type, op1, NULL);
+
+ vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
+ ir->operands[0]->type->vector_elements,
+ ir->operands[0]->type->matrix_columns);
+
+ if (ir->operands[0]->type->base_type == GLSL_TYPE_INT)
+ op0 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[0], NULL);
+ else
+ op0 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[0], NULL);
+
+ op0 = new(ir) ir_expression(ir_binop_mul, vec_type, op0, op1);
+
+ ir->operation = ir_unop_f2i;
+ ir->operands[0] = op0;
+ ir->operands[1] = NULL;
+ }
+
+ this->progress = true;
+}
+
+void
+lower_instructions_visitor::exp_to_exp2(ir_expression *ir)
+{
+ ir_constant *log2_e = new(ir) ir_constant(float(M_LOG2E));
+
+ ir->operation = ir_unop_exp2;
+ ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[0]->type,
+ ir->operands[0], log2_e);
+ this->progress = true;
+}
+
+void
+lower_instructions_visitor::pow_to_exp2(ir_expression *ir)
+{
+ ir_expression *const log2_x =
+ new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type,
+ ir->operands[0]);
+
+ ir->operation = ir_unop_exp2;
+ ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[1]->type,
+ ir->operands[1], log2_x);
+ ir->operands[1] = NULL;
+ this->progress = true;
+}
+
+void
+lower_instructions_visitor::log_to_log2(ir_expression *ir)
+{
+ ir->operation = ir_binop_mul;
+ ir->operands[0] = new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type,
+ ir->operands[0], NULL);
+ ir->operands[1] = new(ir) ir_constant(float(1.0 / M_LOG2E));
+ this->progress = true;
+}
+
+void
+lower_instructions_visitor::mod_to_fract(ir_expression *ir)
+{
+ ir_variable *temp = new(ir) ir_variable(ir->operands[1]->type, "mod_b",
+ ir_var_temporary);
+ this->base_ir->insert_before(temp);
+
+ ir_assignment *const assign =
+ new(ir) ir_assignment(new(ir) ir_dereference_variable(temp),
+ ir->operands[1], NULL);
+
+ this->base_ir->insert_before(assign);
+
+ ir_expression *const div_expr =
+ new(ir) ir_expression(ir_binop_div, ir->operands[0]->type,
+ ir->operands[0],
+ new(ir) ir_dereference_variable(temp));
+
+ /* Don't generate new IR that would need to be lowered in an additional
+ * pass.
+ */
+ if (lowering(DIV_TO_MUL_RCP))
+ div_to_mul_rcp(div_expr);
+
+ ir_rvalue *expr = new(ir) ir_expression(ir_unop_fract,
+ ir->operands[0]->type,
+ div_expr,
+ NULL);
+
+ ir->operation = ir_binop_mul;
+ ir->operands[0] = new(ir) ir_dereference_variable(temp);
+ ir->operands[1] = expr;
+ this->progress = true;
+}
+
+ir_visitor_status
+lower_instructions_visitor::visit_leave(ir_expression *ir)
+{
+ switch (ir->operation) {
+ case ir_binop_sub:
+ if (lowering(SUB_TO_ADD_NEG))
+ sub_to_add_neg(ir);
+ break;
+
+ case ir_binop_div:
+ if (lowering(DIV_TO_MUL_RCP))
+ div_to_mul_rcp(ir);
+ break;
+
+ case ir_unop_exp:
+ if (lowering(EXP_TO_EXP2))
+ exp_to_exp2(ir);
+ break;
+
+ case ir_unop_log:
+ if (lowering(LOG_TO_LOG2))
+ log_to_log2(ir);
+ break;
+
+ case ir_binop_mod:
+ if (lowering(MOD_TO_FRACT))
+ mod_to_fract(ir);
+ break;
+
+ case ir_binop_pow:
+ if (lowering(POW_TO_EXP2))
+ pow_to_exp2(ir);
+ break;
+
+ default:
+ return visit_continue;
+ }
+
+ return visit_continue;
+}