/* * Copyright © 2014 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. */ #include "intel_nir.h" #include "brw_nir.h" #include "compiler/glsl_types.h" #include "compiler/nir/nir_builder.h" /* * Returns the minimum number of vec4 (as_vec4 == true) or dvec4 (as_vec4 == * false) elements needed to pack a type. */ static int type_size_xvec4(const struct glsl_type *type, bool as_vec4, bool bindless) { unsigned int i; int size; switch (type->base_type) { case GLSL_TYPE_UINT: case GLSL_TYPE_INT: case GLSL_TYPE_FLOAT: case GLSL_TYPE_FLOAT16: case GLSL_TYPE_BOOL: case GLSL_TYPE_DOUBLE: case GLSL_TYPE_UINT16: case GLSL_TYPE_INT16: case GLSL_TYPE_UINT8: case GLSL_TYPE_INT8: case GLSL_TYPE_UINT64: case GLSL_TYPE_INT64: if (glsl_type_is_matrix(type)) { const glsl_type *col_type = glsl_get_column_type(type); unsigned col_slots = (as_vec4 && glsl_type_is_dual_slot(col_type)) ? 2 : 1; return type->matrix_columns * col_slots; } else { /* Regardless of size of vector, it gets a vec4. This is bad * packing for things like floats, but otherwise arrays become a * mess. Hopefully a later pass over the code can pack scalars * down if appropriate. */ return (as_vec4 && glsl_type_is_dual_slot(type)) ? 2 : 1; } case GLSL_TYPE_ARRAY: assert(type->length > 0); return type_size_xvec4(type->fields.array, as_vec4, bindless) * type->length; case GLSL_TYPE_STRUCT: case GLSL_TYPE_INTERFACE: size = 0; for (i = 0; i < type->length; i++) { size += type_size_xvec4(type->fields.structure[i].type, as_vec4, bindless); } return size; case GLSL_TYPE_SUBROUTINE: return 1; case GLSL_TYPE_SAMPLER: case GLSL_TYPE_TEXTURE: /* Samplers and textures take up no register space, since they're baked * in at link time. */ return bindless ? 1 : 0; case GLSL_TYPE_ATOMIC_UINT: return 0; case GLSL_TYPE_IMAGE: return bindless ? 1 : 0; case GLSL_TYPE_VOID: case GLSL_TYPE_ERROR: case GLSL_TYPE_COOPERATIVE_MATRIX: unreachable("not reached"); } return 0; } /** * Returns the minimum number of vec4 elements needed to pack a type. * * For simple types, it will return 1 (a single vec4); for matrices, the * number of columns; for array and struct, the sum of the vec4_size of * each of its elements; and for sampler and atomic, zero. * * This method is useful to calculate how much register space is needed to * store a particular type. */ int type_size_vec4(const struct glsl_type *type, bool bindless) { return type_size_xvec4(type, true, bindless); } /** * Returns the minimum number of dvec4 elements needed to pack a type. * * For simple types, it will return 1 (a single dvec4); for matrices, the * number of columns; for array and struct, the sum of the dvec4_size of * each of its elements; and for sampler and atomic, zero. * * This method is useful to calculate how much register space is needed to * store a particular type. * * Measuring double-precision vertex inputs as dvec4 is required because * ARB_vertex_attrib_64bit states that these uses the same number of locations * than the single-precision version. That is, two consecutives dvec4 would be * located in location "x" and location "x+1", not "x+2". * * In order to map vec4/dvec4 vertex inputs in the proper ATTRs, * remap_vs_attrs() will take in account both the location and also if the * type fits in one or two vec4 slots. */ int type_size_dvec4(const struct glsl_type *type, bool bindless) { return type_size_xvec4(type, false, bindless); } static bool remap_tess_levels(nir_builder *b, nir_intrinsic_instr *intr, enum tess_primitive_mode _primitive_mode) { const int location = nir_intrinsic_base(intr); const unsigned component = nir_intrinsic_component(intr); bool out_of_bounds = false; bool write = !nir_intrinsic_infos[intr->intrinsic].has_dest; unsigned mask = write ? nir_intrinsic_write_mask(intr) : 0; nir_def *src = NULL, *dest = NULL; if (write) { assert(intr->num_components == intr->src[0].ssa->num_components); } else { assert(intr->num_components == intr->def.num_components); } if (location == VARYING_SLOT_TESS_LEVEL_INNER) { b->cursor = write ? nir_before_instr(&intr->instr) : nir_after_instr(&intr->instr); switch (_primitive_mode) { case TESS_PRIMITIVE_QUADS: /* gl_TessLevelInner[0..1] lives at DWords 3-2 (reversed). */ nir_intrinsic_set_base(intr, 0); if (write) { assert(intr->src[0].ssa->num_components == 2); intr->num_components = 4; nir_def *undef = nir_undef(b, 1, 32); nir_def *x = nir_channel(b, intr->src[0].ssa, 0); nir_def *y = nir_channel(b, intr->src[0].ssa, 1); src = nir_vec4(b, undef, undef, y, x); mask = !!(mask & WRITEMASK_X) << 3 | !!(mask & WRITEMASK_Y) << 2; } else if (intr->def.num_components > 1) { assert(intr->def.num_components == 2); intr->num_components = 4; intr->def.num_components = 4; unsigned wz[2] = { 3, 2 }; dest = nir_swizzle(b, &intr->def, wz, 2); } else { nir_intrinsic_set_component(intr, 3 - component); } break; case TESS_PRIMITIVE_TRIANGLES: /* gl_TessLevelInner[0] lives at DWord 4. */ nir_intrinsic_set_base(intr, 1); mask &= WRITEMASK_X; out_of_bounds = component > 0; break; case TESS_PRIMITIVE_ISOLINES: out_of_bounds = true; break; default: unreachable("Bogus tessellation domain"); } } else if (location == VARYING_SLOT_TESS_LEVEL_OUTER) { b->cursor = write ? nir_before_instr(&intr->instr) : nir_after_instr(&intr->instr); nir_intrinsic_set_base(intr, 1); switch (_primitive_mode) { case TESS_PRIMITIVE_QUADS: case TESS_PRIMITIVE_TRIANGLES: /* Quads: gl_TessLevelOuter[0..3] lives at DWords 7-4 (reversed). * Triangles: gl_TessLevelOuter[0..2] lives at DWords 7-5 (reversed). */ if (write) { assert(intr->src[0].ssa->num_components == 4); unsigned wzyx[4] = { 3, 2, 1, 0 }; src = nir_swizzle(b, intr->src[0].ssa, wzyx, 4); mask = !!(mask & WRITEMASK_X) << 3 | !!(mask & WRITEMASK_Y) << 2 | !!(mask & WRITEMASK_Z) << 1 | !!(mask & WRITEMASK_W) << 0; /* Don't overwrite the inner factor at DWord 4 for triangles */ if (_primitive_mode == TESS_PRIMITIVE_TRIANGLES) mask &= ~WRITEMASK_X; } else if (intr->def.num_components > 1) { assert(intr->def.num_components == 4); unsigned wzyx[4] = { 3, 2, 1, 0 }; dest = nir_swizzle(b, &intr->def, wzyx, 4); } else { nir_intrinsic_set_component(intr, 3 - component); out_of_bounds = component == 3 && _primitive_mode == TESS_PRIMITIVE_TRIANGLES; } break; case TESS_PRIMITIVE_ISOLINES: /* gl_TessLevelOuter[0..1] lives at DWords 6-7 (in order). */ if (write) { assert(intr->src[0].ssa->num_components == 4); nir_def *undef = nir_undef(b, 1, 32); nir_def *x = nir_channel(b, intr->src[0].ssa, 0); nir_def *y = nir_channel(b, intr->src[0].ssa, 1); src = nir_vec4(b, undef, undef, x, y); mask = !!(mask & WRITEMASK_X) << 2 | !!(mask & WRITEMASK_Y) << 3; } else { nir_intrinsic_set_component(intr, 2 + component); out_of_bounds = component > 1; } break; default: unreachable("Bogus tessellation domain"); } } else { return false; } if (out_of_bounds) { if (!write) nir_def_rewrite_uses(&intr->def, nir_undef(b, 1, 32)); nir_instr_remove(&intr->instr); } else if (write) { nir_intrinsic_set_write_mask(intr, mask); if (src) { nir_src_rewrite(&intr->src[0], src); } } else if (dest) { nir_def_rewrite_uses_after(&intr->def, dest, dest->parent_instr); } return true; } static bool is_input(nir_intrinsic_instr *intrin) { return intrin->intrinsic == nir_intrinsic_load_input || intrin->intrinsic == nir_intrinsic_load_per_primitive_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 intel_vue_map *vue_map, enum tess_primitive_mode tes_primitive_mode) { nir_foreach_instr_safe(instr, block) { if (instr->type != nir_instr_type_intrinsic) continue; nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); gl_shader_stage stage = b->shader->info.stage; if ((stage == MESA_SHADER_TESS_CTRL && is_output(intrin)) || (stage == MESA_SHADER_TESS_EVAL && is_input(intrin))) { if (remap_tess_levels(b, intrin, tes_primitive_mode)) continue; int vue_slot = vue_map->varying_to_slot[intrin->const_index[0]]; assert(vue_slot != -1); intrin->const_index[0] = vue_slot; nir_src *vertex = nir_get_io_arrayed_index_src(intrin); if (vertex) { if (nir_src_is_const(*vertex)) { intrin->const_index[0] += nir_src_as_uint(*vertex) * vue_map->num_per_vertex_slots; } else { b->cursor = nir_before_instr(&intrin->instr); /* Multiply by the number of per-vertex slots. */ nir_def *vertex_offset = nir_imul(b, vertex->ssa, nir_imm_int(b, vue_map->num_per_vertex_slots)); /* Add it to the existing offset */ nir_src *offset = nir_get_io_offset_src(intrin); nir_def *total_offset = nir_iadd(b, vertex_offset, offset->ssa); nir_src_rewrite(offset, total_offset); } } } } return true; } void brw_nir_lower_vs_inputs(nir_shader *nir) { /* Start with the location of the variable's base. */ nir_foreach_shader_in_variable(var, nir) var->data.driver_location = var->data.location; /* Now use nir_lower_io to walk dereference chains. Attribute arrays are * loaded as one vec4 or dvec4 per element (or matrix column), depending on * whether it is a double-precision type or not. */ nir_lower_io(nir, nir_var_shader_in, type_size_vec4, nir_lower_io_lower_64bit_to_32); /* This pass needs actual constants */ nir_opt_constant_folding(nir); nir_io_add_const_offset_to_base(nir, nir_var_shader_in); /* The last step is to remap VERT_ATTRIB_* to actual registers */ /* Whether or not we have any system generated values. gl_DrawID is not * included here as it lives in its own vec4. */ const bool has_sgvs = BITSET_TEST(nir->info.system_values_read, SYSTEM_VALUE_FIRST_VERTEX) || BITSET_TEST(nir->info.system_values_read, SYSTEM_VALUE_BASE_INSTANCE) || BITSET_TEST(nir->info.system_values_read, SYSTEM_VALUE_VERTEX_ID_ZERO_BASE) || BITSET_TEST(nir->info.system_values_read, SYSTEM_VALUE_INSTANCE_ID); const unsigned num_inputs = util_bitcount64(nir->info.inputs_read); nir_foreach_function_impl(impl, nir) { nir_builder b = nir_builder_create(impl); nir_foreach_block(block, impl) { nir_foreach_instr_safe(instr, block) { if (instr->type != nir_instr_type_intrinsic) continue; nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); switch (intrin->intrinsic) { case nir_intrinsic_load_first_vertex: case nir_intrinsic_load_base_instance: case nir_intrinsic_load_vertex_id_zero_base: case nir_intrinsic_load_instance_id: case nir_intrinsic_load_is_indexed_draw: case nir_intrinsic_load_draw_id: { b.cursor = nir_after_instr(&intrin->instr); /* gl_VertexID and friends are stored by the VF as the last * vertex element. We convert them to load_input intrinsics at * the right location. */ nir_intrinsic_instr *load = nir_intrinsic_instr_create(nir, nir_intrinsic_load_input); load->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0)); nir_intrinsic_set_base(load, num_inputs); switch (intrin->intrinsic) { case nir_intrinsic_load_first_vertex: nir_intrinsic_set_component(load, 0); break; case nir_intrinsic_load_base_instance: nir_intrinsic_set_component(load, 1); break; case nir_intrinsic_load_vertex_id_zero_base: nir_intrinsic_set_component(load, 2); break; case nir_intrinsic_load_instance_id: nir_intrinsic_set_component(load, 3); break; case nir_intrinsic_load_draw_id: case nir_intrinsic_load_is_indexed_draw: /* gl_DrawID and IsIndexedDraw are stored right after * gl_VertexID and friends if any of them exist. */ nir_intrinsic_set_base(load, num_inputs + has_sgvs); if (intrin->intrinsic == nir_intrinsic_load_draw_id) nir_intrinsic_set_component(load, 0); else nir_intrinsic_set_component(load, 1); break; default: unreachable("Invalid system value intrinsic"); } load->num_components = 1; nir_def_init(&load->instr, &load->def, 1, 32); nir_builder_instr_insert(&b, &load->instr); nir_def_replace(&intrin->def, &load->def); break; } case nir_intrinsic_load_input: { /* Attributes come in a contiguous block, ordered by their * gl_vert_attrib value. That means we can compute the slot * number for an attribute by masking out the enabled attributes * before it and counting the bits. */ int attr = nir_intrinsic_base(intrin); int slot = util_bitcount64(nir->info.inputs_read & BITFIELD64_MASK(attr)); nir_intrinsic_set_base(intrin, slot); break; } default: break; /* Nothing to do */ } } } } } void brw_nir_lower_vue_inputs(nir_shader *nir, const struct intel_vue_map *vue_map) { nir_foreach_shader_in_variable(var, nir) var->data.driver_location = var->data.location; /* Inputs are stored in vec4 slots, so use type_size_vec4(). */ nir_lower_io(nir, nir_var_shader_in, type_size_vec4, nir_lower_io_lower_64bit_to_32); /* This pass needs actual constants */ nir_opt_constant_folding(nir); nir_io_add_const_offset_to_base(nir, nir_var_shader_in); nir_foreach_function_impl(impl, nir) { nir_foreach_block(block, impl) { nir_foreach_instr(instr, block) { if (instr->type != nir_instr_type_intrinsic) continue; nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); if (intrin->intrinsic == nir_intrinsic_load_input || intrin->intrinsic == nir_intrinsic_load_per_vertex_input) { /* Offset 0 is the VUE header, which contains * VARYING_SLOT_LAYER [.y], VARYING_SLOT_VIEWPORT [.z], and * VARYING_SLOT_PSIZ [.w]. */ int varying = nir_intrinsic_base(intrin); int vue_slot; switch (varying) { case VARYING_SLOT_PSIZ: nir_intrinsic_set_base(intrin, 0); nir_intrinsic_set_component(intrin, 3); break; default: vue_slot = vue_map->varying_to_slot[varying]; assert(vue_slot != -1); nir_intrinsic_set_base(intrin, vue_slot); break; } } } } } } void brw_nir_lower_tes_inputs(nir_shader *nir, const struct intel_vue_map *vue_map) { nir_foreach_shader_in_variable(var, nir) var->data.driver_location = var->data.location; nir_lower_io(nir, nir_var_shader_in, type_size_vec4, nir_lower_io_lower_64bit_to_32); /* This pass needs actual constants */ nir_opt_constant_folding(nir); nir_io_add_const_offset_to_base(nir, nir_var_shader_in); nir_foreach_function_impl(impl, nir) { nir_builder b = nir_builder_create(impl); nir_foreach_block(block, impl) { remap_patch_urb_offsets(block, &b, vue_map, nir->info.tess._primitive_mode); } } } static bool lower_barycentric_per_sample(nir_builder *b, nir_intrinsic_instr *intrin, UNUSED void *cb_data) { if (intrin->intrinsic != nir_intrinsic_load_barycentric_pixel && intrin->intrinsic != nir_intrinsic_load_barycentric_centroid) return false; b->cursor = nir_before_instr(&intrin->instr); nir_def *centroid = nir_load_barycentric(b, nir_intrinsic_load_barycentric_sample, nir_intrinsic_interp_mode(intrin)); nir_def_replace(&intrin->def, centroid); return true; } /** * Convert interpolateAtOffset() offsets from [-0.5, +0.5] floating point * offsets to integer [-8, +7] offsets (in units of 1/16th of a pixel). * * We clamp to +7/16 on the upper end of the range, since +0.5 isn't * representable in a S0.4 value; a naive conversion would give us -8/16, * which is the opposite of what was intended. * * This is allowed by GL_ARB_gpu_shader5's quantization rules: * * "Not all values of may be supported; x and y offsets may * be rounded to fixed-point values with the number of fraction bits * given by the implementation-dependent constant * FRAGMENT_INTERPOLATION_OFFSET_BITS." */ static bool lower_barycentric_at_offset(nir_builder *b, nir_intrinsic_instr *intrin, void *data) { if (intrin->intrinsic != nir_intrinsic_load_barycentric_at_offset) return false; b->cursor = nir_before_instr(&intrin->instr); assert(intrin->src[0].ssa); nir_def *offset = nir_imin(b, nir_imm_int(b, 7), nir_f2i32(b, nir_fmul_imm(b, intrin->src[0].ssa, 16))); nir_src_rewrite(&intrin->src[0], offset); return true; } void brw_nir_lower_fs_inputs(nir_shader *nir, const struct intel_device_info *devinfo, const struct brw_wm_prog_key *key) { nir_foreach_shader_in_variable(var, nir) { var->data.driver_location = var->data.location; /* Apply default interpolation mode. * * Everything defaults to smooth except for the legacy GL color * built-in variables, which might be flat depending on API state. */ if (var->data.interpolation == INTERP_MODE_NONE) { const bool flat = key->flat_shade && (var->data.location == VARYING_SLOT_COL0 || var->data.location == VARYING_SLOT_COL1); var->data.interpolation = flat ? INTERP_MODE_FLAT : INTERP_MODE_SMOOTH; } } nir_lower_io(nir, nir_var_shader_in, type_size_vec4, nir_lower_io_lower_64bit_to_32); if (devinfo->ver >= 11) nir_lower_interpolation(nir, ~0); if (key->multisample_fbo == BRW_NEVER) { nir_lower_single_sampled(nir); } else if (key->persample_interp == BRW_ALWAYS) { nir_shader_intrinsics_pass(nir, lower_barycentric_per_sample, nir_metadata_control_flow, NULL); } if (devinfo->ver < 20) nir_shader_intrinsics_pass(nir, lower_barycentric_at_offset, nir_metadata_control_flow, NULL); /* This pass needs actual constants */ nir_opt_constant_folding(nir); nir_io_add_const_offset_to_base(nir, nir_var_shader_in); } void brw_nir_lower_vue_outputs(nir_shader *nir) { nir_foreach_shader_out_variable(var, nir) { var->data.driver_location = var->data.location; } nir_lower_io(nir, nir_var_shader_out, type_size_vec4, nir_lower_io_lower_64bit_to_32); } void brw_nir_lower_tcs_outputs(nir_shader *nir, const struct intel_vue_map *vue_map, enum tess_primitive_mode tes_primitive_mode) { nir_foreach_shader_out_variable(var, nir) { var->data.driver_location = var->data.location; } nir_lower_io(nir, nir_var_shader_out, type_size_vec4, nir_lower_io_lower_64bit_to_32); /* This pass needs actual constants */ nir_opt_constant_folding(nir); nir_io_add_const_offset_to_base(nir, nir_var_shader_out); nir_foreach_function_impl(impl, nir) { nir_builder b = nir_builder_create(impl); nir_foreach_block(block, impl) { remap_patch_urb_offsets(block, &b, vue_map, tes_primitive_mode); } } } void brw_nir_lower_fs_outputs(nir_shader *nir) { nir_foreach_shader_out_variable(var, nir) { var->data.driver_location = SET_FIELD(var->data.index, BRW_NIR_FRAG_OUTPUT_INDEX) | SET_FIELD(var->data.location, BRW_NIR_FRAG_OUTPUT_LOCATION); } nir_lower_io(nir, nir_var_shader_out, type_size_dvec4, 0); } static bool tag_speculative_access(nir_builder *b, nir_intrinsic_instr *intrin, void *unused) { if (intrin->intrinsic == nir_intrinsic_load_ubo && brw_nir_ubo_surface_index_is_pushable(intrin->src[0])) { nir_intrinsic_set_access(intrin, ACCESS_CAN_SPECULATE | nir_intrinsic_access(intrin)); return true; } return false; } static bool brw_nir_tag_speculative_access(nir_shader *nir) { return nir_shader_intrinsics_pass(nir, tag_speculative_access, nir_metadata_all, NULL); } #define OPT(pass, ...) ({ \ bool this_progress = false; \ NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__); \ if (this_progress) \ progress = true; \ this_progress; \ }) #define LOOP_OPT(pass, ...) ({ \ const unsigned long this_line = __LINE__; \ bool this_progress = false; \ if (opt_line == this_line) \ break; \ NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__); \ if (this_progress) { \ progress = true; \ opt_line = this_line; \ } \ this_progress; \ }) #define LOOP_OPT_NOT_IDEMPOTENT(pass, ...) ({ \ bool this_progress = false; \ NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__); \ if (this_progress) { \ progress = true; \ opt_line = 0; \ } \ this_progress; \ }) void brw_nir_optimize(nir_shader *nir, const struct intel_device_info *devinfo) { bool progress; unsigned lower_flrp = (nir->options->lower_flrp16 ? 16 : 0) | (nir->options->lower_flrp32 ? 32 : 0) | (nir->options->lower_flrp64 ? 64 : 0); unsigned long opt_line = 0; do { progress = false; /* This pass is causing problems with types used by OpenCL : * https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13955 * * Running with it disabled made no difference in the resulting assembly * code. */ if (nir->info.stage != MESA_SHADER_KERNEL) LOOP_OPT(nir_split_array_vars, nir_var_function_temp); LOOP_OPT(nir_shrink_vec_array_vars, nir_var_function_temp); LOOP_OPT(nir_opt_deref); if (LOOP_OPT(nir_opt_memcpy)) LOOP_OPT(nir_split_var_copies); LOOP_OPT(nir_lower_vars_to_ssa); if (!nir->info.var_copies_lowered) { /* Only run this pass if nir_lower_var_copies was not called * yet. That would lower away any copy_deref instructions and we * don't want to introduce any more. */ LOOP_OPT(nir_opt_find_array_copies); } LOOP_OPT(nir_opt_copy_prop_vars); LOOP_OPT(nir_opt_dead_write_vars); LOOP_OPT(nir_opt_combine_stores, nir_var_all); LOOP_OPT(nir_opt_ray_queries); LOOP_OPT(nir_opt_ray_query_ranges); LOOP_OPT(nir_lower_alu_to_scalar, NULL, NULL); LOOP_OPT(nir_copy_prop); LOOP_OPT(nir_lower_phis_to_scalar, false); LOOP_OPT(nir_copy_prop); LOOP_OPT(nir_opt_dce); LOOP_OPT(nir_opt_cse); LOOP_OPT(nir_opt_combine_stores, nir_var_all); /* Passing 0 to the peephole select pass causes it to convert * if-statements that contain only move instructions in the branches * regardless of the count. * * Passing 1 to the peephole select pass causes it to convert * if-statements that contain at most a single ALU instruction (total) * in both branches. Before Gfx6, some math instructions were * prohibitively expensive and the results of compare operations need an * extra resolve step. For these reasons, this pass is more harmful * than good on those platforms. * * For indirect loads of uniforms (push constants), we assume that array * indices will nearly always be in bounds and the cost of the load is * low. Therefore there shouldn't be a performance benefit to avoid it. */ LOOP_OPT(nir_opt_peephole_select, 0, true, false); LOOP_OPT(nir_opt_peephole_select, 8, true, true); LOOP_OPT(nir_opt_intrinsics); LOOP_OPT(nir_opt_idiv_const, 32); LOOP_OPT_NOT_IDEMPOTENT(nir_opt_algebraic); LOOP_OPT(nir_opt_generate_bfi); LOOP_OPT(nir_opt_reassociate_bfi); LOOP_OPT(nir_lower_constant_convert_alu_types); LOOP_OPT(nir_opt_constant_folding); if (lower_flrp != 0) { if (LOOP_OPT(nir_lower_flrp, lower_flrp, false /* always_precise */)) { LOOP_OPT(nir_opt_constant_folding); } /* Nothing should rematerialize any flrps, so we only need to do this * lowering once. */ lower_flrp = 0; } LOOP_OPT(nir_opt_dead_cf); if (LOOP_OPT(nir_opt_loop)) { /* If nir_opt_loop makes progress, then we need to clean * things up if we want any hope of nir_opt_if or nir_opt_loop_unroll * to make progress. */ LOOP_OPT(nir_copy_prop); LOOP_OPT(nir_opt_dce); } LOOP_OPT_NOT_IDEMPOTENT(nir_opt_if, nir_opt_if_optimize_phi_true_false); LOOP_OPT(nir_opt_conditional_discard); if (nir->options->max_unroll_iterations != 0) { LOOP_OPT_NOT_IDEMPOTENT(nir_opt_loop_unroll); } LOOP_OPT(nir_opt_remove_phis); LOOP_OPT(nir_opt_gcm, false); LOOP_OPT(nir_opt_undef); LOOP_OPT(nir_lower_pack); } while (progress); /* Workaround Gfxbench unused local sampler variable which will trigger an * assert in the opt_large_constants pass. */ OPT(nir_remove_dead_variables, nir_var_function_temp, NULL); } static unsigned lower_bit_size_callback(const nir_instr *instr, UNUSED void *data) { switch (instr->type) { case nir_instr_type_alu: { nir_alu_instr *alu = nir_instr_as_alu(instr); switch (alu->op) { case nir_op_bit_count: case nir_op_ufind_msb: case nir_op_ifind_msb: case nir_op_find_lsb: /* These are handled specially because the destination is always * 32-bit and so the bit size of the instruction is given by the * source. */ return alu->src[0].src.ssa->bit_size >= 32 ? 0 : 32; default: break; } if (alu->def.bit_size >= 32) return 0; /* Note: nir_op_iabs and nir_op_ineg are not lowered here because the * 8-bit ABS or NEG instruction should eventually get copy propagated * into the MOV that does the type conversion. This results in far * fewer MOV instructions. */ switch (alu->op) { case nir_op_idiv: case nir_op_imod: case nir_op_irem: case nir_op_udiv: case nir_op_umod: case nir_op_fceil: case nir_op_ffloor: case nir_op_ffract: case nir_op_fround_even: case nir_op_ftrunc: return 32; case nir_op_frcp: case nir_op_frsq: case nir_op_fsqrt: case nir_op_fpow: case nir_op_fexp2: case nir_op_flog2: case nir_op_fsin: case nir_op_fcos: return 0; case nir_op_isign: assert(!"Should have been lowered by nir_opt_algebraic."); return 0; default: if (nir_op_infos[alu->op].num_inputs >= 2 && alu->def.bit_size == 8) return 16; if (nir_alu_instr_is_comparison(alu) && alu->src[0].src.ssa->bit_size == 8) return 16; return 0; } break; } case nir_instr_type_intrinsic: { nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); switch (intrin->intrinsic) { case nir_intrinsic_read_invocation: case nir_intrinsic_read_first_invocation: case nir_intrinsic_vote_feq: case nir_intrinsic_vote_ieq: case nir_intrinsic_shuffle: case nir_intrinsic_shuffle_xor: case nir_intrinsic_shuffle_up: case nir_intrinsic_shuffle_down: case nir_intrinsic_quad_broadcast: case nir_intrinsic_quad_swap_horizontal: case nir_intrinsic_quad_swap_vertical: case nir_intrinsic_quad_swap_diagonal: if (intrin->src[0].ssa->bit_size == 8) return 16; return 0; case nir_intrinsic_reduce: case nir_intrinsic_inclusive_scan: case nir_intrinsic_exclusive_scan: /* There are a couple of register region issues that make things * complicated for 8-bit types: * * 1. Only raw moves are allowed to write to a packed 8-bit * destination. * 2. If we use a strided destination, the efficient way to do * scan operations ends up using strides that are too big to * encode in an instruction. * * To get around these issues, we just do all 8-bit scan operations * in 16 bits. It's actually fewer instructions than what we'd have * to do if we were trying to do it in native 8-bit types and the * results are the same once we truncate to 8 bits at the end. */ if (intrin->def.bit_size == 8) return 16; return 0; default: return 0; } break; } case nir_instr_type_phi: { nir_phi_instr *phi = nir_instr_as_phi(instr); if (phi->def.bit_size == 8) return 16; return 0; } default: return 0; } } /* On gfx12.5+, if the offsets are not both constant and in the {-8,7} range, * we will have nir_lower_tex() lower the source offset by returning true from * this filter function. */ static bool lower_xehp_tg4_offset_filter(const nir_instr *instr, UNUSED const void *data) { if (instr->type != nir_instr_type_tex) return false; nir_tex_instr *tex = nir_instr_as_tex(instr); if (tex->op != nir_texop_tg4) return false; int offset_index = nir_tex_instr_src_index(tex, nir_tex_src_offset); if (offset_index < 0) return false; if (!nir_src_is_const(tex->src[offset_index].src)) return true; int64_t offset_x = nir_src_comp_as_int(tex->src[offset_index].src, 0); int64_t offset_y = nir_src_comp_as_int(tex->src[offset_index].src, 1); return offset_x < -8 || offset_x > 7 || offset_y < -8 || offset_y > 7; } /* Does some simple lowering and runs the standard suite of optimizations * * This is intended to be called more-or-less directly after you get the * shader out of GLSL or some other source. While it is geared towards i965, * it is not at all generator-specific. */ void brw_preprocess_nir(const struct brw_compiler *compiler, nir_shader *nir, const struct brw_nir_compiler_opts *opts) { const struct intel_device_info *devinfo = compiler->devinfo; UNUSED bool progress; /* Written by OPT */ nir_validate_ssa_dominance(nir, "before brw_preprocess_nir"); OPT(nir_lower_frexp); OPT(nir_lower_alu_to_scalar, NULL, NULL); struct nir_opt_16bit_tex_image_options options = { .rounding_mode = nir_rounding_mode_undef, .opt_tex_dest_types = nir_type_float | nir_type_int | nir_type_uint, }; OPT(nir_opt_16bit_tex_image, &options); if (nir->info.stage == MESA_SHADER_GEOMETRY) OPT(nir_lower_gs_intrinsics, 0); /* See also brw_nir_trig_workarounds.py */ if (compiler->precise_trig && !(devinfo->ver >= 10 || devinfo->platform == INTEL_PLATFORM_KBL)) OPT(brw_nir_apply_trig_workarounds); /* This workaround existing for performance reasons. Since it requires not * setting RENDER_SURFACE_STATE::SurfaceArray when the array length is 1, * we're loosing the HW robustness feature in that case. * * So when robust image access is enabled, just avoid the workaround. */ if (intel_needs_workaround(devinfo, 1806565034) && !opts->robust_image_access) OPT(intel_nir_clamp_image_1d_2d_array_sizes); const struct intel_nir_lower_texture_opts intel_tex_options = { .combined_lod_or_bias_and_offset = compiler->devinfo->ver >= 20, }; OPT(intel_nir_lower_texture, &intel_tex_options); const nir_lower_tex_options tex_options = { .lower_txp = ~0, .lower_txf_offset = true, .lower_rect_offset = true, .lower_txd_cube_map = true, /* For below, See bspec 45942, "Enable new message layout for cube array" */ .lower_txd_3d = devinfo->verx10 >= 125, .lower_txd_array = devinfo->verx10 >= 125, .lower_txb_shadow_clamp = true, .lower_txd_shadow_clamp = true, .lower_txd_offset_clamp = true, .lower_tg4_offsets = true, .lower_txs_lod = true, /* Wa_14012320009 */ .lower_offset_filter = devinfo->verx10 >= 125 ? lower_xehp_tg4_offset_filter : NULL, .lower_invalid_implicit_lod = true, }; /* In the case where TG4 coords are lowered to offsets and we have a * lower_xehp_tg4_offset_filter lowering those offsets further, we need to * rerun the pass because the instructions inserted by the first lowering * are not visible during that first pass. */ if (OPT(nir_lower_tex, &tex_options)) { OPT(intel_nir_lower_texture, &intel_tex_options); OPT(nir_lower_tex, &tex_options); } OPT(nir_normalize_cubemap_coords); OPT(nir_lower_global_vars_to_local); OPT(nir_split_var_copies); OPT(nir_split_struct_vars, nir_var_function_temp); brw_nir_optimize(nir, devinfo); OPT(nir_lower_doubles, opts->softfp64, nir->options->lower_doubles_options); if (OPT(nir_lower_int64_float_conversions)) { OPT(nir_opt_algebraic); OPT(nir_lower_doubles, opts->softfp64, nir->options->lower_doubles_options); } OPT(nir_lower_bit_size, lower_bit_size_callback, (void *)compiler); /* Lower a bunch of stuff */ OPT(nir_lower_var_copies); /* This needs to be run after the first optimization pass but before we * lower indirect derefs away */ OPT(nir_opt_large_constants, NULL, 32); OPT(nir_lower_load_const_to_scalar); OPT(nir_lower_system_values); nir_lower_compute_system_values_options lower_csv_options = { .has_base_workgroup_id = nir->info.stage == MESA_SHADER_COMPUTE, }; OPT(nir_lower_compute_system_values, &lower_csv_options); const nir_lower_subgroups_options subgroups_options = { .ballot_bit_size = 32, .ballot_components = 1, .lower_to_scalar = true, .lower_relative_shuffle = true, .lower_quad_broadcast_dynamic = true, .lower_elect = true, .lower_inverse_ballot = true, .lower_rotate_to_shuffle = true, }; OPT(nir_lower_subgroups, &subgroups_options); nir_variable_mode indirect_mask = brw_nir_no_indirect_mask(compiler, nir->info.stage); OPT(nir_lower_indirect_derefs, indirect_mask, UINT32_MAX); /* Even in cases where we can handle indirect temporaries via scratch, we * it can still be expensive. Lower indirects on small arrays to * conditional load/stores. * * The threshold of 16 was chosen semi-arbitrarily. The idea is that an * indirect on an array of 16 elements is about 30 instructions at which * point, you may be better off doing a send. With a SIMD8 program, 16 * floats is 1/8 of the entire register file. Any array larger than that * is likely to cause pressure issues. Also, this value is sufficiently * high that the benchmarks known to suffer from large temporary array * issues are helped but nothing else in shader-db is hurt except for maybe * that one kerbal space program shader. */ if (!(indirect_mask & nir_var_function_temp)) OPT(nir_lower_indirect_derefs, nir_var_function_temp, 16); /* Lower array derefs of vectors for SSBO and UBO loads. For both UBOs and * SSBOs, our back-end is capable of loading an entire vec4 at a time and * we would like to take advantage of that whenever possible regardless of * whether or not the app gives us full loads. This should allow the * optimizer to combine UBO and SSBO load operations and save us some send * messages. */ OPT(nir_lower_array_deref_of_vec, nir_var_mem_ubo | nir_var_mem_ssbo, NULL, nir_lower_direct_array_deref_of_vec_load); /* Clamp load_per_vertex_input of the TCS stage so that we do not generate * loads reading out of bounds. We can do this here because we called * nir_lower_system_values above. */ if (nir->info.stage == MESA_SHADER_TESS_CTRL && compiler->use_tcs_multi_patch) OPT(intel_nir_clamp_per_vertex_loads); /* Get rid of split copies */ brw_nir_optimize(nir, devinfo); } static bool brw_nir_zero_inputs_instr(struct nir_builder *b, nir_intrinsic_instr *intrin, void *data) { if (intrin->intrinsic != nir_intrinsic_load_deref) return false; nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]); if (!nir_deref_mode_is(deref, nir_var_shader_in)) return false; if (deref->deref_type != nir_deref_type_var) return false; nir_variable *var = deref->var; uint64_t zero_inputs = *(uint64_t *)data; if (!(BITFIELD64_BIT(var->data.location) & zero_inputs)) return false; b->cursor = nir_before_instr(&intrin->instr); nir_def *zero = nir_imm_zero(b, 1, 32); nir_def_replace(&intrin->def, zero); return true; } static bool brw_nir_zero_inputs(nir_shader *shader, uint64_t *zero_inputs) { return nir_shader_intrinsics_pass(shader, brw_nir_zero_inputs_instr, nir_metadata_control_flow, zero_inputs); } /* Code for Wa_18019110168 may have created input/output variables beyond * VARYING_SLOT_MAX and removed uses of variables below VARYING_SLOT_MAX. * Clean it up, so they all stay below VARYING_SLOT_MAX. */ static void brw_mesh_compact_io(nir_shader *mesh, nir_shader *frag) { gl_varying_slot mapping[VARYING_SLOT_MAX] = {0, }; gl_varying_slot cur = VARYING_SLOT_VAR0; bool compact = false; nir_foreach_shader_out_variable(var, mesh) { gl_varying_slot location = var->data.location; if (location < VARYING_SLOT_VAR0) continue; assert(location < ARRAY_SIZE(mapping)); const struct glsl_type *type = var->type; if (nir_is_arrayed_io(var, MESA_SHADER_MESH) || var->data.per_view) { assert(glsl_type_is_array(type)); type = glsl_get_array_element(type); } if (mapping[location]) continue; unsigned num_slots = glsl_count_attribute_slots(type, false); compact |= location + num_slots > VARYING_SLOT_MAX; mapping[location] = cur; cur += num_slots; } if (!compact) return; /* The rest of this function should be hit only for Wa_18019110168. */ nir_foreach_shader_out_variable(var, mesh) { gl_varying_slot location = var->data.location; if (location < VARYING_SLOT_VAR0) continue; location = mapping[location]; if (location == 0) continue; var->data.location = location; } nir_foreach_shader_in_variable(var, frag) { gl_varying_slot location = var->data.location; if (location < VARYING_SLOT_VAR0) continue; location = mapping[location]; if (location == 0) continue; var->data.location = location; } nir_shader_gather_info(mesh, nir_shader_get_entrypoint(mesh)); nir_shader_gather_info(frag, nir_shader_get_entrypoint(frag)); if (should_print_nir(mesh)) { printf("%s\n", __func__); nir_print_shader(mesh, stdout); } if (should_print_nir(frag)) { printf("%s\n", __func__); nir_print_shader(frag, stdout); } } void brw_nir_link_shaders(const struct brw_compiler *compiler, nir_shader *producer, nir_shader *consumer) { const struct intel_device_info *devinfo = compiler->devinfo; if (producer->info.stage == MESA_SHADER_MESH && consumer->info.stage == MESA_SHADER_FRAGMENT) { uint64_t fs_inputs = 0, ms_outputs = 0; /* gl_MeshPerPrimitiveEXT[].gl_ViewportIndex, gl_PrimitiveID and gl_Layer * are per primitive, but fragment shader does not have them marked as * such. Add the annotation here. */ nir_foreach_shader_in_variable(var, consumer) { fs_inputs |= BITFIELD64_BIT(var->data.location); switch (var->data.location) { case VARYING_SLOT_LAYER: case VARYING_SLOT_PRIMITIVE_ID: case VARYING_SLOT_VIEWPORT: var->data.per_primitive = 1; break; default: continue; } } nir_foreach_shader_out_variable(var, producer) ms_outputs |= BITFIELD64_BIT(var->data.location); uint64_t zero_inputs = ~ms_outputs & fs_inputs; zero_inputs &= BITFIELD64_BIT(VARYING_SLOT_LAYER) | BITFIELD64_BIT(VARYING_SLOT_VIEWPORT); if (zero_inputs) NIR_PASS(_, consumer, brw_nir_zero_inputs, &zero_inputs); } nir_lower_io_arrays_to_elements(producer, consumer); nir_validate_shader(producer, "after nir_lower_io_arrays_to_elements"); nir_validate_shader(consumer, "after nir_lower_io_arrays_to_elements"); NIR_PASS(_, producer, nir_lower_io_to_scalar_early, nir_var_shader_out); NIR_PASS(_, consumer, nir_lower_io_to_scalar_early, nir_var_shader_in); brw_nir_optimize(producer, devinfo); brw_nir_optimize(consumer, devinfo); if (nir_link_opt_varyings(producer, consumer)) brw_nir_optimize(consumer, devinfo); NIR_PASS(_, producer, nir_remove_dead_variables, nir_var_shader_out, NULL); NIR_PASS(_, consumer, nir_remove_dead_variables, nir_var_shader_in, NULL); if (nir_remove_unused_varyings(producer, consumer)) { if (should_print_nir(producer)) { printf("nir_remove_unused_varyings\n"); nir_print_shader(producer, stdout); } if (should_print_nir(consumer)) { printf("nir_remove_unused_varyings\n"); nir_print_shader(consumer, stdout); } NIR_PASS(_, producer, nir_lower_global_vars_to_local); NIR_PASS(_, consumer, nir_lower_global_vars_to_local); /* The backend might not be able to handle indirects on * temporaries so we need to lower indirects on any of the * varyings we have demoted here. */ NIR_PASS(_, producer, nir_lower_indirect_derefs, brw_nir_no_indirect_mask(compiler, producer->info.stage), UINT32_MAX); NIR_PASS(_, consumer, nir_lower_indirect_derefs, brw_nir_no_indirect_mask(compiler, consumer->info.stage), UINT32_MAX); brw_nir_optimize(producer, devinfo); brw_nir_optimize(consumer, devinfo); if (producer->info.stage == MESA_SHADER_MESH && consumer->info.stage == MESA_SHADER_FRAGMENT) { brw_mesh_compact_io(producer, consumer); } } NIR_PASS(_, producer, nir_lower_io_to_vector, nir_var_shader_out); if (producer->info.stage == MESA_SHADER_TESS_CTRL && producer->options->vectorize_tess_levels) NIR_PASS_V(producer, nir_vectorize_tess_levels); NIR_PASS(_, producer, nir_opt_combine_stores, nir_var_shader_out); NIR_PASS(_, consumer, nir_lower_io_to_vector, nir_var_shader_in); if (producer->info.stage != MESA_SHADER_TESS_CTRL && producer->info.stage != MESA_SHADER_MESH && producer->info.stage != MESA_SHADER_TASK) { /* Calling lower_io_to_vector creates output variable writes with * write-masks. On non-TCS outputs, the back-end can't handle it and we * need to call nir_lower_io_to_temporaries to get rid of them. This, * in turn, creates temporary variables and extra copy_deref intrinsics * that we need to clean up. * * Note Mesh/Task don't support I/O as temporaries (I/O is shared * between whole workgroup, possibly using multiple HW threads). For * those write-mask in output is handled by I/O lowering. */ NIR_PASS_V(producer, nir_lower_io_to_temporaries, nir_shader_get_entrypoint(producer), true, false); NIR_PASS(_, producer, nir_lower_global_vars_to_local); NIR_PASS(_, producer, nir_split_var_copies); NIR_PASS(_, producer, nir_lower_var_copies); } if (producer->info.stage == MESA_SHADER_TASK && consumer->info.stage == MESA_SHADER_MESH) { for (unsigned i = 0; i < 3; ++i) assert(producer->info.mesh.ts_mesh_dispatch_dimensions[i] <= UINT16_MAX); nir_lower_compute_system_values_options options = { .lower_workgroup_id_to_index = true, .num_workgroups[0] = producer->info.mesh.ts_mesh_dispatch_dimensions[0], .num_workgroups[1] = producer->info.mesh.ts_mesh_dispatch_dimensions[1], .num_workgroups[2] = producer->info.mesh.ts_mesh_dispatch_dimensions[2], /* nir_lower_idiv generates expensive code */ .shortcut_1d_workgroup_id = compiler->devinfo->verx10 >= 125, }; NIR_PASS(_, consumer, nir_lower_compute_system_values, &options); } } bool brw_nir_should_vectorize_mem(unsigned align_mul, unsigned align_offset, unsigned bit_size, unsigned num_components, nir_intrinsic_instr *low, nir_intrinsic_instr *high, void *data) { /* Don't combine things to generate 64-bit loads/stores. We have to split * those back into 32-bit ones anyway and UBO loads aren't split in NIR so * we don't want to make a mess for the back-end. */ if (bit_size > 32) return false; if (low->intrinsic == nir_intrinsic_load_ubo_uniform_block_intel || low->intrinsic == nir_intrinsic_load_ssbo_uniform_block_intel || low->intrinsic == nir_intrinsic_load_shared_uniform_block_intel || low->intrinsic == nir_intrinsic_load_global_constant_uniform_block_intel) { if (num_components > 4) { if (!util_is_power_of_two_nonzero(num_components)) return false; if (bit_size != 32) return false; if (num_components > 32) return false; } } else { /* We can handle at most a vec4 right now. Anything bigger would get * immediately split by brw_nir_lower_mem_access_bit_sizes anyway. */ if (num_components > 4) return false; } uint32_t align; if (align_offset) align = 1 << (ffs(align_offset) - 1); else align = align_mul; if (align < bit_size / 8) return false; return true; } static bool combine_all_memory_barriers(nir_intrinsic_instr *a, nir_intrinsic_instr *b, void *data) { /* Combine control barriers with identical memory semantics. This prevents * the second barrier generating a spurious, identical fence message as the * first barrier. */ if (nir_intrinsic_memory_modes(a) == nir_intrinsic_memory_modes(b) && nir_intrinsic_memory_semantics(a) == nir_intrinsic_memory_semantics(b) && nir_intrinsic_memory_scope(a) == nir_intrinsic_memory_scope(b)) { nir_intrinsic_set_execution_scope(a, MAX2(nir_intrinsic_execution_scope(a), nir_intrinsic_execution_scope(b))); return true; } /* Only combine pure memory barriers */ if ((nir_intrinsic_execution_scope(a) != SCOPE_NONE) || (nir_intrinsic_execution_scope(b) != SCOPE_NONE)) return false; /* Translation to backend IR will get rid of modes we don't care about, so * no harm in always combining them. * * TODO: While HW has only ACQUIRE|RELEASE fences, we could improve the * scheduling so that it can take advantage of the different semantics. */ nir_intrinsic_set_memory_modes(a, nir_intrinsic_memory_modes(a) | nir_intrinsic_memory_modes(b)); nir_intrinsic_set_memory_semantics(a, nir_intrinsic_memory_semantics(a) | nir_intrinsic_memory_semantics(b)); nir_intrinsic_set_memory_scope(a, MAX2(nir_intrinsic_memory_scope(a), nir_intrinsic_memory_scope(b))); return true; } static nir_mem_access_size_align get_mem_access_size_align(nir_intrinsic_op intrin, uint8_t bytes, uint8_t bit_size, uint32_t align_mul, uint32_t align_offset, bool offset_is_const, const void *cb_data) { const uint32_t align = nir_combined_align(align_mul, align_offset); switch (intrin) { case nir_intrinsic_load_ssbo: case nir_intrinsic_load_shared: case nir_intrinsic_load_scratch: /* The offset is constant so we can use a 32-bit load and just shift it * around as needed. */ if (align < 4 && offset_is_const) { assert(util_is_power_of_two_nonzero(align_mul) && align_mul >= 4); const unsigned pad = align_offset % 4; const unsigned comps32 = MIN2(DIV_ROUND_UP(bytes + pad, 4), 4); return (nir_mem_access_size_align) { .bit_size = 32, .num_components = comps32, .align = 4, }; } break; case nir_intrinsic_load_task_payload: if (bytes < 4 || align < 4) { return (nir_mem_access_size_align) { .bit_size = 32, .num_components = 1, .align = 4, }; } break; default: break; } const bool is_load = nir_intrinsic_infos[intrin].has_dest; const bool is_scratch = intrin == nir_intrinsic_load_scratch || intrin == nir_intrinsic_store_scratch; if (align < 4 || bytes < 4) { /* Choose a byte, word, or dword */ bytes = MIN2(bytes, 4); if (bytes == 3) bytes = is_load ? 4 : 2; if (is_scratch) { /* The way scratch address swizzling works in the back-end, it * happens at a DWORD granularity so we can't have a single load * or store cross a DWORD boundary. */ if ((align_offset % 4) + bytes > MIN2(align_mul, 4)) bytes = MIN2(align_mul, 4) - (align_offset % 4); /* Must be a power of two */ if (bytes == 3) bytes = 2; } return (nir_mem_access_size_align) { .bit_size = bytes * 8, .num_components = 1, .align = 1, }; } else { bytes = MIN2(bytes, 16); return (nir_mem_access_size_align) { .bit_size = 32, .num_components = is_scratch ? 1 : is_load ? DIV_ROUND_UP(bytes, 4) : bytes / 4, .align = 4, }; } } static void brw_vectorize_lower_mem_access(nir_shader *nir, const struct brw_compiler *compiler, enum brw_robustness_flags robust_flags) { bool progress = false; nir_load_store_vectorize_options options = { .modes = nir_var_mem_ubo | nir_var_mem_ssbo | nir_var_mem_global | nir_var_mem_shared | nir_var_mem_task_payload, .callback = brw_nir_should_vectorize_mem, .robust_modes = (nir_variable_mode)0, }; if (robust_flags & BRW_ROBUSTNESS_UBO) options.robust_modes |= nir_var_mem_ubo | nir_var_mem_global; if (robust_flags & BRW_ROBUSTNESS_SSBO) options.robust_modes |= nir_var_mem_ssbo | nir_var_mem_global; OPT(nir_opt_load_store_vectorize, &options); /* Required for nir_divergence_analysis() */ OPT(nir_convert_to_lcssa, true, true); /* When HW supports block loads, using the divergence analysis, try * to find uniform SSBO loads and turn them into block loads. * * Rerun the vectorizer after that to make the largest possible block * loads. * * This is a win on 2 fronts : * - fewer send messages * - reduced register pressure */ nir_divergence_analysis(nir); if (OPT(intel_nir_blockify_uniform_loads, compiler->devinfo)) OPT(nir_opt_load_store_vectorize, &options); OPT(nir_opt_remove_phis); nir_lower_mem_access_bit_sizes_options mem_access_options = { .modes = nir_var_mem_ssbo | nir_var_mem_constant | nir_var_mem_task_payload | nir_var_shader_temp | nir_var_function_temp | nir_var_mem_global | nir_var_mem_shared, .callback = get_mem_access_size_align, }; OPT(nir_lower_mem_access_bit_sizes, &mem_access_options); while (progress) { progress = false; OPT(nir_lower_pack); OPT(nir_copy_prop); OPT(nir_opt_dce); OPT(nir_opt_cse); OPT(nir_opt_algebraic); OPT(nir_opt_constant_folding); } } static bool nir_shader_has_local_variables(const nir_shader *nir) { nir_foreach_function_impl(impl, nir) { if (!exec_list_is_empty(&impl->locals)) return true; } return false; } /* Prepare the given shader for codegen * * This function is intended to be called right before going into the actual * backend and is highly backend-specific. Also, once this function has been * called on a shader, it will no longer be in SSA form so most optimizations * will not work. */ void brw_postprocess_nir(nir_shader *nir, const struct brw_compiler *compiler, bool debug_enabled, enum brw_robustness_flags robust_flags) { const struct intel_device_info *devinfo = compiler->devinfo; UNUSED bool progress; /* Written by OPT */ OPT(intel_nir_lower_sparse_intrinsics); OPT(nir_lower_bit_size, lower_bit_size_callback, (void *)compiler); OPT(nir_opt_combine_barriers, combine_all_memory_barriers, NULL); OPT(intel_nir_lower_printf); do { progress = false; OPT(nir_opt_algebraic_before_ffma); } while (progress); if (devinfo->verx10 >= 125) { /* Lower integer division by constants before nir_lower_idiv. */ OPT(nir_opt_idiv_const, 32); const nir_lower_idiv_options options = { .allow_fp16 = false }; OPT(nir_lower_idiv, &options); } if (gl_shader_stage_can_set_fragment_shading_rate(nir->info.stage)) NIR_PASS(_, nir, intel_nir_lower_shading_rate_output); OPT(brw_nir_tag_speculative_access); brw_nir_optimize(nir, devinfo); if (nir_shader_has_local_variables(nir)) { OPT(nir_lower_vars_to_explicit_types, nir_var_function_temp, glsl_get_natural_size_align_bytes); OPT(nir_lower_explicit_io, nir_var_function_temp, nir_address_format_32bit_offset); brw_nir_optimize(nir, devinfo); } brw_vectorize_lower_mem_access(nir, compiler, robust_flags); /* Potentially perform this optimization pass twice because it can create * additional opportunities for itself. */ if (OPT(nir_opt_algebraic_before_lower_int64)) OPT(nir_opt_algebraic_before_lower_int64); if (OPT(nir_lower_int64)) brw_nir_optimize(nir, devinfo); /* Try and fuse multiply-adds, if successful, run shrink_vectors to * avoid peephole_ffma to generate things like this : * vec16 ssa_0 = ... * vec16 ssa_1 = fneg ssa_0 * vec1 ssa_2 = ffma ssa_1, ... * * We want this instead : * vec16 ssa_0 = ... * vec1 ssa_1 = fneg ssa_0.x * vec1 ssa_2 = ffma ssa_1, ... */ if (OPT(intel_nir_opt_peephole_ffma)) OPT(nir_opt_shrink_vectors, false); OPT(intel_nir_opt_peephole_imul32x16); if (OPT(nir_opt_comparison_pre)) { OPT(nir_copy_prop); OPT(nir_opt_dce); OPT(nir_opt_cse); /* Do the select peepehole again. nir_opt_comparison_pre (combined with * the other optimization passes) will have removed at least one * instruction from one of the branches of the if-statement, so now it * might be under the threshold of conversion to bcsel. */ OPT(nir_opt_peephole_select, 0, false, false); OPT(nir_opt_peephole_select, 1, false, true); } do { progress = false; OPT(brw_nir_opt_fsat); OPT(nir_opt_algebraic_late); OPT(brw_nir_lower_fsign); if (progress) { OPT(nir_opt_constant_folding); OPT(nir_copy_prop); OPT(nir_opt_dce); OPT(nir_opt_cse); } } while (progress); if (OPT(nir_lower_fp16_casts, nir_lower_fp16_split_fp64)) { if (OPT(nir_lower_int64)) { brw_nir_optimize(nir, devinfo); } } OPT(nir_lower_alu_to_scalar, NULL, NULL); while (OPT(nir_opt_algebraic_distribute_src_mods)) { OPT(nir_opt_constant_folding); OPT(nir_copy_prop); OPT(nir_opt_dce); OPT(nir_opt_cse); } OPT(nir_copy_prop); OPT(nir_opt_dce); OPT(nir_opt_move, nir_move_comparisons); OPT(nir_opt_dead_cf); bool divergence_analysis_dirty = false; NIR_PASS(_, nir, nir_convert_to_lcssa, true, true); NIR_PASS_V(nir, nir_divergence_analysis); static const nir_lower_subgroups_options subgroups_options = { .ballot_bit_size = 32, .ballot_components = 1, .lower_elect = true, .lower_subgroup_masks = true, }; if (OPT(nir_opt_uniform_atomics, false)) { OPT(nir_lower_subgroups, &subgroups_options); OPT(nir_opt_algebraic_before_lower_int64); if (OPT(nir_lower_int64)) brw_nir_optimize(nir, devinfo); divergence_analysis_dirty = true; } /* nir_opt_uniform_subgroup can create some operations (e.g., * load_subgroup_lt_mask) that need to be lowered again. */ if (OPT(nir_opt_uniform_subgroup, &subgroups_options)) { /* Some of the optimizations can generate 64-bit integer multiplication * that must be lowered. */ OPT(nir_lower_int64); /* Even if nir_lower_int64 did not make progress, re-run the main * optimization loop. nir_opt_uniform_subgroup may have made some things * that previously appeared divergent be marked as convergent. This * allows the elimination of some loops over, say, a TXF instruction * with a non-uniform texture handle. */ brw_nir_optimize(nir, devinfo); OPT(nir_lower_subgroups, &subgroups_options); } /* Run intel_nir_lower_conversions only after the last tiem * brw_nir_optimize is called. Various optimizations invoked there can * rematerialize the conversions that the lowering pass eliminates. */ OPT(intel_nir_lower_conversions); /* Do this only after the last opt_gcm. GCM will undo this lowering. */ if (nir->info.stage == MESA_SHADER_FRAGMENT) { if (divergence_analysis_dirty) { NIR_PASS(_, nir, nir_convert_to_lcssa, true, true); NIR_PASS_V(nir, nir_divergence_analysis); } OPT(intel_nir_lower_non_uniform_barycentric_at_sample); } /* Clean up LCSSA phis */ OPT(nir_opt_remove_phis); OPT(nir_lower_bool_to_int32); OPT(nir_copy_prop); OPT(nir_opt_dce); OPT(nir_lower_locals_to_regs, 32); if (unlikely(debug_enabled)) { /* Re-index SSA defs so we print more sensible numbers. */ nir_foreach_function_impl(impl, nir) { nir_index_ssa_defs(impl); } fprintf(stderr, "NIR (SSA form) for %s shader:\n", _mesa_shader_stage_to_string(nir->info.stage)); nir_print_shader(nir, stderr); } nir_validate_ssa_dominance(nir, "before nir_convert_from_ssa"); /* Rerun the divergence analysis before convert_from_ssa as this pass has * some assert on consistent divergence flags. */ NIR_PASS(_, nir, nir_convert_to_lcssa, true, true); NIR_PASS_V(nir, nir_divergence_analysis); OPT(nir_convert_from_ssa, true); OPT(nir_opt_dce); if (OPT(nir_opt_rematerialize_compares)) OPT(nir_opt_dce); /* The mesh stages require this pass to be called at the last minute, * but if anything is done by it, it will also constant fold, and that * undoes the work done by nir_trivialize_registers, so call it right * before that one instead. */ if (nir->info.stage == MESA_SHADER_MESH || nir->info.stage == MESA_SHADER_TASK) brw_nir_adjust_payload(nir); nir_trivialize_registers(nir); nir_sweep(nir); if (unlikely(debug_enabled)) { fprintf(stderr, "NIR (final form) for %s shader:\n", _mesa_shader_stage_to_string(nir->info.stage)); nir_print_shader(nir, stderr); } } static unsigned get_subgroup_size(const struct shader_info *info, unsigned max_subgroup_size) { switch (info->subgroup_size) { case SUBGROUP_SIZE_API_CONSTANT: /* We have to use the global constant size. */ return BRW_SUBGROUP_SIZE; case SUBGROUP_SIZE_UNIFORM: /* It has to be uniform across all invocations but can vary per stage * if we want. This gives us a bit more freedom. * * For compute, brw_nir_apply_key is called per-dispatch-width so this * is the actual subgroup size and not a maximum. However, we only * invoke one size of any given compute shader so it's still guaranteed * to be uniform across invocations. */ return max_subgroup_size; case SUBGROUP_SIZE_VARYING: /* The subgroup size is allowed to be fully varying. For geometry * stages, we know it's always 8 which is max_subgroup_size so we can * return that. For compute, brw_nir_apply_key is called once per * dispatch-width so max_subgroup_size is the real subgroup size. * * For fragment, we return 0 and let it fall through to the back-end * compiler. This means we can't optimize based on subgroup size but * that's a risk the client took when it asked for a varying subgroup * size. */ return info->stage == MESA_SHADER_FRAGMENT ? 0 : max_subgroup_size; case SUBGROUP_SIZE_REQUIRE_4: unreachable("Unsupported subgroup size type"); case SUBGROUP_SIZE_REQUIRE_8: case SUBGROUP_SIZE_REQUIRE_16: case SUBGROUP_SIZE_REQUIRE_32: assert(gl_shader_stage_uses_workgroup(info->stage) || (info->stage >= MESA_SHADER_RAYGEN && info->stage <= MESA_SHADER_CALLABLE)); /* These enum values are expressly chosen to be equal to the subgroup * size that they require. */ return info->subgroup_size; case SUBGROUP_SIZE_FULL_SUBGROUPS: case SUBGROUP_SIZE_REQUIRE_64: case SUBGROUP_SIZE_REQUIRE_128: break; } unreachable("Invalid subgroup size type"); } unsigned brw_nir_api_subgroup_size(const nir_shader *nir, unsigned hw_subgroup_size) { return get_subgroup_size(&nir->info, hw_subgroup_size); } void brw_nir_apply_key(nir_shader *nir, const struct brw_compiler *compiler, const struct brw_base_prog_key *key, unsigned max_subgroup_size) { bool progress = false; nir_lower_tex_options nir_tex_opts = { .lower_txd_clamp_bindless_sampler = true, .lower_txd_clamp_if_sampler_index_not_lt_16 = true, .lower_invalid_implicit_lod = true, .lower_index_to_offset = true, }; OPT(nir_lower_tex, &nir_tex_opts); const struct intel_nir_lower_texture_opts tex_opts = { .combined_lod_and_array_index = compiler->devinfo->ver >= 20, }; OPT(intel_nir_lower_texture, &tex_opts); const nir_lower_subgroups_options subgroups_options = { .subgroup_size = get_subgroup_size(&nir->info, max_subgroup_size), .ballot_bit_size = 32, .ballot_components = 1, .lower_subgroup_masks = true, }; OPT(nir_lower_subgroups, &subgroups_options); if (key->limit_trig_input_range) OPT(brw_nir_limit_trig_input_range_workaround); if (progress) { brw_nir_optimize(nir, compiler->devinfo); } } enum brw_conditional_mod brw_cmod_for_nir_comparison(nir_op op) { switch (op) { case nir_op_flt: case nir_op_flt32: case nir_op_ilt: case nir_op_ilt32: case nir_op_ult: case nir_op_ult32: return BRW_CONDITIONAL_L; case nir_op_fge: case nir_op_fge32: case nir_op_ige: case nir_op_ige32: case nir_op_uge: case nir_op_uge32: return BRW_CONDITIONAL_GE; case nir_op_feq: case nir_op_feq32: case nir_op_ieq: case nir_op_ieq32: case nir_op_b32all_fequal2: case nir_op_b32all_iequal2: case nir_op_b32all_fequal3: case nir_op_b32all_iequal3: case nir_op_b32all_fequal4: case nir_op_b32all_iequal4: return BRW_CONDITIONAL_Z; case nir_op_fneu: case nir_op_fneu32: case nir_op_ine: case nir_op_ine32: case nir_op_b32any_fnequal2: case nir_op_b32any_inequal2: case nir_op_b32any_fnequal3: case nir_op_b32any_inequal3: case nir_op_b32any_fnequal4: case nir_op_b32any_inequal4: return BRW_CONDITIONAL_NZ; default: unreachable("Unsupported NIR comparison op"); } } enum lsc_opcode lsc_op_for_nir_intrinsic(const nir_intrinsic_instr *intrin) { switch (intrin->intrinsic) { case nir_intrinsic_load_ssbo: case nir_intrinsic_load_shared: case nir_intrinsic_load_global: case nir_intrinsic_load_global_block_intel: case nir_intrinsic_load_global_constant: case nir_intrinsic_load_global_constant_uniform_block_intel: case nir_intrinsic_load_shared_block_intel: case nir_intrinsic_load_shared_uniform_block_intel: case nir_intrinsic_load_ssbo_block_intel: case nir_intrinsic_load_ssbo_uniform_block_intel: case nir_intrinsic_load_ubo_uniform_block_intel: case nir_intrinsic_load_scratch: return LSC_OP_LOAD; case nir_intrinsic_store_ssbo: case nir_intrinsic_store_shared: case nir_intrinsic_store_global: case nir_intrinsic_store_global_block_intel: case nir_intrinsic_store_shared_block_intel: case nir_intrinsic_store_ssbo_block_intel: case nir_intrinsic_store_scratch: return LSC_OP_STORE; case nir_intrinsic_image_load: case nir_intrinsic_bindless_image_load: return LSC_OP_LOAD_CMASK; case nir_intrinsic_image_store: case nir_intrinsic_bindless_image_store: return LSC_OP_STORE_CMASK; default: assert(nir_intrinsic_has_atomic_op(intrin)); break; } switch (nir_intrinsic_atomic_op(intrin)) { case nir_atomic_op_iadd: { unsigned src_idx; switch (intrin->intrinsic) { case nir_intrinsic_image_atomic: case nir_intrinsic_bindless_image_atomic: src_idx = 3; break; case nir_intrinsic_ssbo_atomic: src_idx = 2; break; case nir_intrinsic_shared_atomic: case nir_intrinsic_global_atomic: src_idx = 1; break; default: unreachable("Invalid add atomic opcode"); } if (nir_src_is_const(intrin->src[src_idx])) { int64_t add_val = nir_src_as_int(intrin->src[src_idx]); if (add_val == 1) return LSC_OP_ATOMIC_INC; else if (add_val == -1) return LSC_OP_ATOMIC_DEC; } return LSC_OP_ATOMIC_ADD; } case nir_atomic_op_imin: return LSC_OP_ATOMIC_MIN; case nir_atomic_op_umin: return LSC_OP_ATOMIC_UMIN; case nir_atomic_op_imax: return LSC_OP_ATOMIC_MAX; case nir_atomic_op_umax: return LSC_OP_ATOMIC_UMAX; case nir_atomic_op_iand: return LSC_OP_ATOMIC_AND; case nir_atomic_op_ior: return LSC_OP_ATOMIC_OR; case nir_atomic_op_ixor: return LSC_OP_ATOMIC_XOR; case nir_atomic_op_xchg: return LSC_OP_ATOMIC_STORE; case nir_atomic_op_cmpxchg: return LSC_OP_ATOMIC_CMPXCHG; case nir_atomic_op_fmin: return LSC_OP_ATOMIC_FMIN; case nir_atomic_op_fmax: return LSC_OP_ATOMIC_FMAX; case nir_atomic_op_fcmpxchg: return LSC_OP_ATOMIC_FCMPXCHG; case nir_atomic_op_fadd: return LSC_OP_ATOMIC_FADD; default: unreachable("Unsupported NIR atomic intrinsic"); } } enum brw_reg_type brw_type_for_nir_type(const struct intel_device_info *devinfo, nir_alu_type type) { switch (type) { case nir_type_uint: case nir_type_uint32: return BRW_TYPE_UD; case nir_type_bool: case nir_type_int: case nir_type_bool32: case nir_type_int32: return BRW_TYPE_D; case nir_type_float: case nir_type_float32: return BRW_TYPE_F; case nir_type_float16: return BRW_TYPE_HF; case nir_type_float64: return BRW_TYPE_DF; case nir_type_int64: return BRW_TYPE_Q; case nir_type_uint64: return BRW_TYPE_UQ; case nir_type_int16: return BRW_TYPE_W; case nir_type_uint16: return BRW_TYPE_UW; case nir_type_int8: return BRW_TYPE_B; case nir_type_uint8: return BRW_TYPE_UB; default: unreachable("unknown type"); } return BRW_TYPE_F; } nir_shader * brw_nir_create_passthrough_tcs(void *mem_ctx, const struct brw_compiler *compiler, const struct brw_tcs_prog_key *key) { assert(key->input_vertices > 0); const nir_shader_compiler_options *options = compiler->nir_options[MESA_SHADER_TESS_CTRL]; uint64_t inputs_read = key->outputs_written & ~(VARYING_BIT_TESS_LEVEL_INNER | VARYING_BIT_TESS_LEVEL_OUTER); unsigned locations[64]; unsigned num_locations = 0; u_foreach_bit64(varying, inputs_read) locations[num_locations++] = varying; nir_shader *nir = nir_create_passthrough_tcs_impl(options, locations, num_locations, key->input_vertices); ralloc_steal(mem_ctx, nir); nir->info.inputs_read = inputs_read; nir->info.tess._primitive_mode = key->_tes_primitive_mode; nir_validate_shader(nir, "in brw_nir_create_passthrough_tcs"); struct brw_nir_compiler_opts opts = {}; brw_preprocess_nir(compiler, nir, &opts); return nir; } nir_def * brw_nir_load_global_const(nir_builder *b, nir_intrinsic_instr *load_uniform, nir_def *base_addr, unsigned off) { assert(load_uniform->intrinsic == nir_intrinsic_load_uniform); unsigned bit_size = load_uniform->def.bit_size; assert(bit_size >= 8 && bit_size % 8 == 0); unsigned byte_size = bit_size / 8; nir_def *sysval; if (nir_src_is_const(load_uniform->src[0])) { uint64_t offset = off + nir_intrinsic_base(load_uniform) + nir_src_as_uint(load_uniform->src[0]); /* Things should be component-aligned. */ assert(offset % byte_size == 0); unsigned suboffset = offset % 64; uint64_t aligned_offset = offset - suboffset; /* Load two just in case we go over a 64B boundary */ nir_def *data[2]; for (unsigned i = 0; i < 2; i++) { nir_def *addr = nir_iadd_imm(b, base_addr, aligned_offset + i * 64); data[i] = nir_load_global_constant_uniform_block_intel( b, 16, 32, addr, .access = ACCESS_CAN_REORDER | ACCESS_NON_WRITEABLE, .align_mul = 64); } sysval = nir_extract_bits(b, data, 2, suboffset * 8, load_uniform->num_components, bit_size); } else { nir_def *offset32 = nir_iadd_imm(b, load_uniform->src[0].ssa, off + nir_intrinsic_base(load_uniform)); nir_def *addr = nir_iadd(b, base_addr, nir_u2u64(b, offset32)); sysval = nir_load_global_constant(b, addr, byte_size, load_uniform->num_components, bit_size); } return sysval; } const struct glsl_type * brw_nir_get_var_type(const struct nir_shader *nir, nir_variable *var) { const struct glsl_type *type = var->interface_type; if (!type) { type = var->type; if (nir_is_arrayed_io(var, nir->info.stage) || var->data.per_view) { assert(glsl_type_is_array(type)); type = glsl_get_array_element(type); } } return type; }