From 63684a9ae7a66f68df1f2c68cd9358e5622122a3 Mon Sep 17 00:00:00 2001 From: Kenneth Graunke Date: Thu, 18 Nov 2010 17:54:07 -0800 Subject: glsl: Combine many instruction lowering passes into one. This should save on the overhead of tree-walking and provide a convenient place to add more instruction lowering in the future. Signed-off-by: Ian Romanick --- src/glsl/lower_instructions.cpp | 262 ++++++++++++++++++++++++++++++++++++++++ 1 file changed, 262 insertions(+) create mode 100644 src/glsl/lower_instructions.cpp (limited to 'src/glsl/lower_instructions.cpp') diff --git a/src/glsl/lower_instructions.cpp b/src/glsl/lower_instructions.cpp new file mode 100644 index 00000000000..d460ba1a97a --- /dev/null +++ b/src/glsl/lower_instructions.cpp @@ -0,0 +1,262 @@ +/* + * 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 + * - 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. + * + * 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_E */ +#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 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(log2f(M_E)); + + 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::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(1.0f / log2f(M_E)); + 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; + + default: + return visit_continue; + } + + return visit_continue; +} -- cgit v1.2.3 From da61afa7388f1ce50ef612b89aba2302a052a3bb Mon Sep 17 00:00:00 2001 From: Ian Romanick Date: Wed, 24 Nov 2010 14:03:57 -0800 Subject: glsl: Use M_LOG2E constant instead of calling log2 --- src/glsl/lower_instructions.cpp | 6 +++--- src/mesa/main/compiler.h | 4 ++++ 2 files changed, 7 insertions(+), 3 deletions(-) (limited to 'src/glsl/lower_instructions.cpp') diff --git a/src/glsl/lower_instructions.cpp b/src/glsl/lower_instructions.cpp index d460ba1a97a..0d9374d73bd 100644 --- a/src/glsl/lower_instructions.cpp +++ b/src/glsl/lower_instructions.cpp @@ -70,7 +70,7 @@ * opportunity to do things like constant fold the (1.0 / op1) easily. */ -#include "main/core.h" /* for M_E */ +#include "main/core.h" /* for M_LOG2E */ #include "glsl_types.h" #include "ir.h" #include "ir_optimization.h" @@ -172,7 +172,7 @@ lower_instructions_visitor::div_to_mul_rcp(ir_expression *ir) void lower_instructions_visitor::exp_to_exp2(ir_expression *ir) { - ir_constant *log2_e = new(ir) ir_constant(log2f(M_E)); + 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, @@ -186,7 +186,7 @@ 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(1.0f / log2f(M_E)); + ir->operands[1] = new(ir) ir_constant(float(1.0 / M_LOG2E)); this->progress = true; } diff --git a/src/mesa/main/compiler.h b/src/mesa/main/compiler.h index 800eb839005..5557a3b5cb5 100644 --- a/src/mesa/main/compiler.h +++ b/src/mesa/main/compiler.h @@ -358,6 +358,10 @@ static INLINE GLuint CPU_TO_LE32(GLuint x) #define M_E (2.7182818284590452354) #endif +#ifndef M_LOG2E +#define M_LOG2E (1.4426950408889634074) +#endif + #ifndef ONE_DIV_LN2 #define ONE_DIV_LN2 (1.442695040888963456) #endif -- cgit v1.2.3 From c4285be9a5bd1adaa89050989374b95a9a601cdc Mon Sep 17 00:00:00 2001 From: Ian Romanick Date: Wed, 24 Nov 2010 22:21:10 -0800 Subject: glsl: Lower ir_binop_pow to a sequence of EXP2 and LOG2 --- src/glsl/ir_optimization.h | 5 +++-- src/glsl/lower_instructions.cpp | 26 ++++++++++++++++++++++++++ src/mesa/main/mtypes.h | 1 + src/mesa/program/ir_to_mesa.cpp | 5 +++-- 4 files changed, 33 insertions(+), 4 deletions(-) (limited to 'src/glsl/lower_instructions.cpp') diff --git a/src/glsl/ir_optimization.h b/src/glsl/ir_optimization.h index 1048ff982e0..f264265f4b1 100644 --- a/src/glsl/ir_optimization.h +++ b/src/glsl/ir_optimization.h @@ -32,8 +32,9 @@ #define SUB_TO_ADD_NEG 0x01 #define DIV_TO_MUL_RCP 0x02 #define EXP_TO_EXP2 0x04 -#define LOG_TO_LOG2 0x08 -#define MOD_TO_FRACT 0x10 +#define POW_TO_EXP2 0x08 +#define LOG_TO_LOG2 0x10 +#define MOD_TO_FRACT 0x20 bool do_common_optimization(exec_list *ir, bool linked, unsigned max_unroll_iterations); diff --git a/src/glsl/lower_instructions.cpp b/src/glsl/lower_instructions.cpp index 0d9374d73bd..a5f61f213d0 100644 --- a/src/glsl/lower_instructions.cpp +++ b/src/glsl/lower_instructions.cpp @@ -33,6 +33,7 @@ * - SUB_TO_ADD_NEG * - DIV_TO_MUL_RCP * - EXP_TO_EXP2 + * - POW_TO_EXP2 * - LOG_TO_LOG2 * - MOD_TO_FRACT * @@ -61,6 +62,11 @@ * 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)) @@ -91,6 +97,7 @@ private: 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 *); }; @@ -180,6 +187,20 @@ lower_instructions_visitor::exp_to_exp2(ir_expression *ir) 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) { @@ -254,6 +275,11 @@ lower_instructions_visitor::visit_leave(ir_expression *ir) mod_to_fract(ir); break; + case ir_binop_pow: + if (lowering(POW_TO_EXP2)) + pow_to_exp2(ir); + break; + default: return visit_continue; } diff --git a/src/mesa/main/mtypes.h b/src/mesa/main/mtypes.h index 80c20e09d9a..82495714f21 100644 --- a/src/mesa/main/mtypes.h +++ b/src/mesa/main/mtypes.h @@ -2197,6 +2197,7 @@ struct gl_shader_compiler_options GLboolean EmitNoCont; /**< Emit CONT opcode? */ GLboolean EmitNoMainReturn; /**< Emit CONT/RET opcodes? */ GLboolean EmitNoNoise; /**< Emit NOISE opcodes? */ + GLboolean EmitNoPow; /**< Emit POW opcodes? */ /** * \name Forms of indirect addressing the driver cannot do. diff --git a/src/mesa/program/ir_to_mesa.cpp b/src/mesa/program/ir_to_mesa.cpp index d9d86b6c29f..b274a961b28 100644 --- a/src/mesa/program/ir_to_mesa.cpp +++ b/src/mesa/program/ir_to_mesa.cpp @@ -2849,8 +2849,9 @@ _mesa_ir_link_shader(struct gl_context *ctx, struct gl_shader_program *prog) /* Lowering */ do_mat_op_to_vec(ir); - lower_instructions(ir, MOD_TO_FRACT | DIV_TO_MUL_RCP | EXP_TO_EXP2 - | LOG_TO_LOG2); + lower_instructions(ir, (MOD_TO_FRACT | DIV_TO_MUL_RCP | EXP_TO_EXP2 + | LOG_TO_LOG2 + | ((options->EmitNoPow) ? POW_TO_EXP2 : 0))); progress = do_lower_jumps(ir, true, true, options->EmitNoMainReturn, options->EmitNoCont, options->EmitNoLoops) || progress; -- cgit v1.2.3