From: Connor Abbott Date: Tue, 14 May 2019 10:10:11 +0000 (+0200) Subject: intel/nir: Extract add_const_offset_to_base X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=6b28808b22848dd5bfcda658f7b1d42f9ee06d73;p=mesa.git intel/nir: Extract add_const_offset_to_base Pretty much every driver using nir_lower_io_to_temporaries followed by nir_lower_io is going to want this. In particular, radv and radeonsi in the next commits. Reviewed-by: Bas Nieuwenhuizen --- diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h index a2b3d393f12..55a39a493e9 100644 --- a/src/compiler/nir/nir.h +++ b/src/compiler/nir/nir.h @@ -3171,6 +3171,8 @@ bool nir_lower_io(nir_shader *shader, int (*type_size)(const struct glsl_type *, bool), nir_lower_io_options); +bool nir_io_add_const_offset_to_base(nir_shader *nir, nir_variable_mode mode); + typedef enum { /** * An address format which is a simple 32-bit global GPU address. diff --git a/src/compiler/nir/nir_lower_io.c b/src/compiler/nir/nir_lower_io.c index a9b3b95ef49..9275803d6f5 100644 --- a/src/compiler/nir/nir_lower_io.c +++ b/src/compiler/nir/nir_lower_io.c @@ -1307,3 +1307,78 @@ nir_build_addr_isub(nir_builder *b, nir_ssa_def *addr0, nir_ssa_def *addr1, unreachable("Invalid address format"); } + +static bool +is_input(nir_intrinsic_instr *intrin) +{ + return intrin->intrinsic == nir_intrinsic_load_input || + intrin->intrinsic == nir_intrinsic_load_per_vertex_input || + intrin->intrinsic == nir_intrinsic_load_interpolated_input || + intrin->intrinsic == nir_intrinsic_load_fs_input_interp_deltas; +} + +static bool +is_output(nir_intrinsic_instr *intrin) +{ + return intrin->intrinsic == nir_intrinsic_load_output || + intrin->intrinsic == nir_intrinsic_load_per_vertex_output || + intrin->intrinsic == nir_intrinsic_store_output || + intrin->intrinsic == nir_intrinsic_store_per_vertex_output; +} + + +/** + * This pass adds constant offsets to instr->const_index[0] for input/output + * intrinsics, and resets the offset source to 0. Non-constant offsets remain + * unchanged - since we don't know what part of a compound variable is + * accessed, we allocate storage for the entire thing. For drivers that use + * nir_lower_io_to_temporaries() before nir_lower_io(), this guarantees that + * the offset source will be 0, so that they don't have to add it in manually. + */ + +static bool +add_const_offset_to_base_block(nir_block *block, nir_builder *b, + nir_variable_mode mode) +{ + bool progress = false; + nir_foreach_instr_safe(instr, block) { + if (instr->type != nir_instr_type_intrinsic) + continue; + + nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); + + if ((mode == nir_var_shader_in && is_input(intrin)) || + (mode == nir_var_shader_out && is_output(intrin))) { + nir_src *offset = nir_get_io_offset_src(intrin); + + if (nir_src_is_const(*offset)) { + intrin->const_index[0] += nir_src_as_uint(*offset); + b->cursor = nir_before_instr(&intrin->instr); + nir_instr_rewrite_src(&intrin->instr, offset, + nir_src_for_ssa(nir_imm_int(b, 0))); + progress = true; + } + } + } + + return progress; +} + +bool +nir_io_add_const_offset_to_base(nir_shader *nir, nir_variable_mode mode) +{ + bool progress = false; + + nir_foreach_function(f, nir) { + if (f->impl) { + nir_builder b; + nir_builder_init(&b, f->impl); + nir_foreach_block(block, f->impl) { + progress |= add_const_offset_to_base_block(block, &b, mode); + } + } + } + + return progress; +} + diff --git a/src/intel/compiler/brw_nir.c b/src/intel/compiler/brw_nir.c index 4dea234dd4b..2d46b87de6c 100644 --- a/src/intel/compiler/brw_nir.c +++ b/src/intel/compiler/brw_nir.c @@ -28,75 +28,6 @@ #include "compiler/nir/nir_builder.h" #include "util/u_math.h" -static bool -is_input(nir_intrinsic_instr *intrin) -{ - return intrin->intrinsic == nir_intrinsic_load_input || - intrin->intrinsic == nir_intrinsic_load_per_vertex_input || - intrin->intrinsic == nir_intrinsic_load_interpolated_input || - intrin->intrinsic == nir_intrinsic_load_fs_input_interp_deltas; -} - -static bool -is_output(nir_intrinsic_instr *intrin) -{ - return intrin->intrinsic == nir_intrinsic_load_output || - intrin->intrinsic == nir_intrinsic_load_per_vertex_output || - intrin->intrinsic == nir_intrinsic_store_output || - intrin->intrinsic == nir_intrinsic_store_per_vertex_output; -} - -/** - * In many cases, we just add the base and offset together, so there's no - * reason to keep them separate. Sometimes, combining them is essential: - * if a shader only accesses part of a compound variable (such as a matrix - * or array), the variable's base may not actually exist in the VUE map. - * - * This pass adds constant offsets to instr->const_index[0], and resets - * the offset source to 0. Non-constant offsets remain unchanged - since - * we don't know what part of a compound variable is accessed, we allocate - * storage for the entire thing. - */ - -static bool -add_const_offset_to_base_block(nir_block *block, nir_builder *b, - nir_variable_mode mode) -{ - nir_foreach_instr_safe(instr, block) { - if (instr->type != nir_instr_type_intrinsic) - continue; - - nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); - - if ((mode == nir_var_shader_in && is_input(intrin)) || - (mode == nir_var_shader_out && is_output(intrin))) { - nir_src *offset = nir_get_io_offset_src(intrin); - - if (nir_src_is_const(*offset)) { - intrin->const_index[0] += nir_src_as_uint(*offset); - b->cursor = nir_before_instr(&intrin->instr); - nir_instr_rewrite_src(&intrin->instr, offset, - nir_src_for_ssa(nir_imm_int(b, 0))); - } - } - } - return true; -} - -static void -add_const_offset_to_base(nir_shader *nir, nir_variable_mode mode) -{ - nir_foreach_function(f, nir) { - if (f->impl) { - nir_builder b; - nir_builder_init(&b, f->impl); - nir_foreach_block(block, f->impl) { - add_const_offset_to_base_block(block, &b, mode); - } - } - } -} - static bool remap_tess_levels(nir_builder *b, nir_intrinsic_instr *intr, GLenum primitive_mode) @@ -152,6 +83,24 @@ remap_tess_levels(nir_builder *b, nir_intrinsic_instr *intr, return true; } +static bool +is_input(nir_intrinsic_instr *intrin) +{ + return intrin->intrinsic == nir_intrinsic_load_input || + intrin->intrinsic == nir_intrinsic_load_per_vertex_input || + intrin->intrinsic == nir_intrinsic_load_interpolated_input; +} + +static bool +is_output(nir_intrinsic_instr *intrin) +{ + return intrin->intrinsic == nir_intrinsic_load_output || + intrin->intrinsic == nir_intrinsic_load_per_vertex_output || + intrin->intrinsic == nir_intrinsic_store_output || + intrin->intrinsic == nir_intrinsic_store_per_vertex_output; +} + + static bool remap_patch_urb_offsets(nir_block *block, nir_builder *b, const struct brw_vue_map *vue_map, @@ -227,7 +176,7 @@ brw_nir_lower_vs_inputs(nir_shader *nir, /* This pass needs actual constants */ nir_opt_constant_folding(nir); - add_const_offset_to_base(nir, nir_var_shader_in); + nir_io_add_const_offset_to_base(nir, nir_var_shader_in); brw_nir_apply_attribute_workarounds(nir, vs_attrib_wa_flags); @@ -350,7 +299,7 @@ brw_nir_lower_vue_inputs(nir_shader *nir, /* This pass needs actual constants */ nir_opt_constant_folding(nir); - add_const_offset_to_base(nir, nir_var_shader_in); + nir_io_add_const_offset_to_base(nir, nir_var_shader_in); nir_foreach_function(function, nir) { if (!function->impl) @@ -401,7 +350,7 @@ brw_nir_lower_tes_inputs(nir_shader *nir, const struct brw_vue_map *vue_map) /* This pass needs actual constants */ nir_opt_constant_folding(nir); - add_const_offset_to_base(nir, nir_var_shader_in); + nir_io_add_const_offset_to_base(nir, nir_var_shader_in); nir_foreach_function(function, nir) { if (function->impl) { @@ -458,7 +407,7 @@ brw_nir_lower_fs_inputs(nir_shader *nir, /* This pass needs actual constants */ nir_opt_constant_folding(nir); - add_const_offset_to_base(nir, nir_var_shader_in); + nir_io_add_const_offset_to_base(nir, nir_var_shader_in); } void @@ -484,7 +433,7 @@ brw_nir_lower_tcs_outputs(nir_shader *nir, const struct brw_vue_map *vue_map, /* This pass needs actual constants */ nir_opt_constant_folding(nir); - add_const_offset_to_base(nir, nir_var_shader_out); + nir_io_add_const_offset_to_base(nir, nir_var_shader_out); nir_foreach_function(function, nir) { if (function->impl) {