From 857f3a68ea7370d1db2ef5faca181a36ada7d49d Mon Sep 17 00:00:00 2001 From: Kenneth Graunke Date: Tue, 8 Apr 2014 18:09:43 -0700 Subject: [PATCH] glsl: Ignore loop-too-large heuristic if there's bad variable indexing. Many shaders use a pattern such as: for (int i = 0; i < NUM_LIGHTS; i++) { ...access a uniform array, or shader input/output array... } where NUM_LIGHTS is a small constant (such as 2, 4, or 8). The expectation is that the compiler will unroll those loops, turning the array access into constant indexing, which is more efficient, and which may enable array splitting and other optimizations. In many cases, our heuristic fails - either there's another tiny nested loop inside, or the estimated number of instructions is just barely beyond the threshold. So, we fail to unroll the loop, leaving the variable indexing in place. Drivers which don't support the particular flavor of variable indexing will call lower_variable_index_to_cond_assign(), which generates piles and piles of immensely inefficient code. We'd like to avoid generating that. This patch detects unsupported forms of variable-indexing in loops, where the array index is a loop induction variable. In that case, it bypasses the loop-too-large heuristic and forces unrolling. Improves performance in various microbenchmarks: Gl32PSBump8 by 47%, Gl32ShMapVsm by 80%, and Gl32ShMapPcf by 27%. No changes in shader-db. v2: Check ir->array for being an array or matrix, rather than the ir_dereference_array itself. v3: Fix and expand statistics in commit message. Signed-off-by: Kenneth Graunke Reviewed-by: Eric Anholt --- src/glsl/loop_unroll.cpp | 61 ++++++++++++++++++++++++++++++++++++++-- 1 file changed, 58 insertions(+), 3 deletions(-) diff --git a/src/glsl/loop_unroll.cpp b/src/glsl/loop_unroll.cpp index 1ce4d5820e5..da532804a4a 100644 --- a/src/glsl/loop_unroll.cpp +++ b/src/glsl/loop_unroll.cpp @@ -63,13 +63,17 @@ is_break(ir_instruction *ir) class loop_unroll_count : public ir_hierarchical_visitor { public: int nodes; + bool unsupported_variable_indexing; /* If there are nested loops, the node count will be inaccurate. */ bool nested_loop; - loop_unroll_count(exec_list *list) + loop_unroll_count(exec_list *list, loop_variable_state *ls, + const struct gl_shader_compiler_options *options) + : ls(ls), options(options) { nodes = 0; nested_loop = false; + unsupported_variable_indexing = false; run(list); } @@ -91,6 +95,54 @@ public: nested_loop = true; return visit_continue; } + + virtual ir_visitor_status visit_enter(ir_dereference_array *ir) + { + /* Check for arrays variably-indexed by a loop induction variable. + * Unrolling the loop may convert that access into constant-indexing. + * + * Many drivers don't support particular kinds of variable indexing, + * and have to resort to using lower_variable_index_to_cond_assign to + * handle it. This results in huge amounts of horrible code, so we'd + * like to avoid that if possible. Here, we just note that it will + * happen. + */ + if ((ir->array->type->is_array() || ir->array->type->is_matrix()) && + !ir->array_index->as_constant()) { + ir_variable *array = ir->array->variable_referenced(); + loop_variable *lv = ls->get(ir->array_index->variable_referenced()); + if (array && lv && lv->is_induction_var()) { + switch (array->data.mode) { + case ir_var_auto: + case ir_var_temporary: + case ir_var_const_in: + case ir_var_function_in: + case ir_var_function_out: + case ir_var_function_inout: + if (options->EmitNoIndirectTemp) + unsupported_variable_indexing = true; + break; + case ir_var_uniform: + if (options->EmitNoIndirectUniform) + unsupported_variable_indexing = true; + break; + case ir_var_shader_in: + if (options->EmitNoIndirectInput) + unsupported_variable_indexing = true; + break; + case ir_var_shader_out: + if (options->EmitNoIndirectOutput) + unsupported_variable_indexing = true; + break; + } + } + } + return visit_continue; + } + +private: + loop_variable_state *ls; + const struct gl_shader_compiler_options *options; }; @@ -257,9 +309,12 @@ loop_unroll_visitor::visit_leave(ir_loop *ir) /* Don't try to unroll nested loops and loops with a huge body. */ - loop_unroll_count count(&ir->body_instructions); + loop_unroll_count count(&ir->body_instructions, ls, options); + + bool loop_too_large = + count.nested_loop || count.nodes * iterations > max_iterations * 5; - if (count.nested_loop || count.nodes * iterations > max_iterations * 5) + if (loop_too_large && !count.unsupported_variable_indexing) return visit_continue; /* Note: the limiting terminator contributes 1 to ls->num_loop_jumps. -- 2.30.2