/* * 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. * * Authors: * Connor Abbott (cwabbott0@gmail.com) * */ #include "nir.h" #include "nir_builder.h" #include "util/u_math.h" #include "util/set.h" struct lower_sysval_state { const nir_lower_compute_system_values_options *options; /* List of intrinsics that have already been lowered and shouldn't be * lowered again. */ struct set *lower_once_list; }; static nir_ssa_def * sanitize_32bit_sysval(nir_builder *b, nir_intrinsic_instr *intrin) { assert(intrin->dest.is_ssa); const unsigned bit_size = intrin->dest.ssa.bit_size; if (bit_size == 32) return NULL; intrin->dest.ssa.bit_size = 32; return nir_u2uN(b, &intrin->dest.ssa, bit_size); } static nir_ssa_def* build_global_group_size(nir_builder *b, unsigned bit_size) { nir_ssa_def *group_size = nir_load_workgroup_size(b); nir_ssa_def *num_workgroups = nir_load_num_workgroups(b, bit_size); return nir_imul(b, nir_u2uN(b, group_size, bit_size), num_workgroups); } static bool lower_system_value_filter(const nir_instr *instr, const void *_state) { return instr->type == nir_instr_type_intrinsic; } static nir_ssa_def * lower_system_value_instr(nir_builder *b, nir_instr *instr, void *_state) { nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); /* All the intrinsics we care about are loads */ if (!nir_intrinsic_infos[intrin->intrinsic].has_dest) return NULL; assert(intrin->dest.is_ssa); const unsigned bit_size = intrin->dest.ssa.bit_size; switch (intrin->intrinsic) { case nir_intrinsic_load_vertex_id: if (b->shader->options->vertex_id_zero_based) { return nir_iadd(b, nir_load_vertex_id_zero_base(b), nir_load_first_vertex(b)); } else { return NULL; } case nir_intrinsic_load_base_vertex: /** * From the OpenGL 4.6 (11.1.3.9 Shader Inputs) specification: * * "gl_BaseVertex holds the integer value passed to the baseVertex * parameter to the command that resulted in the current shader * invocation. In the case where the command has no baseVertex * parameter, the value of gl_BaseVertex is zero." */ if (b->shader->options->lower_base_vertex) { return nir_iand(b, nir_load_is_indexed_draw(b), nir_load_first_vertex(b)); } else { return NULL; } case nir_intrinsic_load_helper_invocation: if (b->shader->options->lower_helper_invocation) { return nir_build_lowered_load_helper_invocation(b); } else { return NULL; } case nir_intrinsic_load_local_invocation_id: case nir_intrinsic_load_local_invocation_index: case nir_intrinsic_load_workgroup_size: return sanitize_32bit_sysval(b, intrin); case nir_intrinsic_load_deref: { nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]); if (!nir_deref_mode_is(deref, nir_var_system_value)) return NULL; nir_ssa_def *column = NULL; if (deref->deref_type != nir_deref_type_var) { /* The only one system values that aren't plane variables are * gl_SampleMask which is always an array of one element and a * couple of ray-tracing intrinsics which are matrices. */ assert(deref->deref_type == nir_deref_type_array); assert(deref->arr.index.is_ssa); column = deref->arr.index.ssa; nir_deref_instr *arr_deref = deref; deref = nir_deref_instr_parent(deref); assert(deref->deref_type == nir_deref_type_var); switch (deref->var->data.location) { case SYSTEM_VALUE_TESS_LEVEL_INNER: case SYSTEM_VALUE_TESS_LEVEL_OUTER: { nir_ssa_def *index = nir_ssa_for_src(b, arr_deref->arr.index, 1); nir_ssa_def *sysval = (deref->var->data.location == SYSTEM_VALUE_TESS_LEVEL_INNER) ? nir_load_tess_level_inner(b) : nir_load_tess_level_outer(b); return nir_vector_extract(b, sysval, index); } case SYSTEM_VALUE_SAMPLE_MASK_IN: case SYSTEM_VALUE_RAY_OBJECT_TO_WORLD: case SYSTEM_VALUE_RAY_WORLD_TO_OBJECT: case SYSTEM_VALUE_MESH_VIEW_INDICES: case SYSTEM_VALUE_RAY_TRIANGLE_VERTEX_POSITIONS: /* These are all single-element arrays in our implementation, and * the sysval load below just drops the 0 array index. */ break; default: unreachable("unsupported system value array deref"); } } nir_variable *var = deref->var; switch (var->data.location) { case SYSTEM_VALUE_INSTANCE_INDEX: return nir_iadd(b, nir_load_instance_id(b), nir_load_base_instance(b)); case SYSTEM_VALUE_SUBGROUP_EQ_MASK: case SYSTEM_VALUE_SUBGROUP_GE_MASK: case SYSTEM_VALUE_SUBGROUP_GT_MASK: case SYSTEM_VALUE_SUBGROUP_LE_MASK: case SYSTEM_VALUE_SUBGROUP_LT_MASK: { nir_intrinsic_op op = nir_intrinsic_from_system_value(var->data.location); nir_intrinsic_instr *load = nir_intrinsic_instr_create(b->shader, op); nir_ssa_dest_init_for_type(&load->instr, &load->dest, var->type); load->num_components = load->dest.ssa.num_components; nir_builder_instr_insert(b, &load->instr); return &load->dest.ssa; } case SYSTEM_VALUE_DEVICE_INDEX: if (b->shader->options->lower_device_index_to_zero) return nir_imm_int(b, 0); break; case SYSTEM_VALUE_GLOBAL_GROUP_SIZE: return build_global_group_size(b, bit_size); case SYSTEM_VALUE_BARYCENTRIC_LINEAR_PIXEL: return nir_load_barycentric(b, nir_intrinsic_load_barycentric_pixel, INTERP_MODE_NOPERSPECTIVE); case SYSTEM_VALUE_BARYCENTRIC_LINEAR_CENTROID: return nir_load_barycentric(b, nir_intrinsic_load_barycentric_centroid, INTERP_MODE_NOPERSPECTIVE); case SYSTEM_VALUE_BARYCENTRIC_LINEAR_SAMPLE: return nir_load_barycentric(b, nir_intrinsic_load_barycentric_sample, INTERP_MODE_NOPERSPECTIVE); case SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL: return nir_load_barycentric(b, nir_intrinsic_load_barycentric_pixel, INTERP_MODE_SMOOTH); case SYSTEM_VALUE_BARYCENTRIC_PERSP_CENTROID: return nir_load_barycentric(b, nir_intrinsic_load_barycentric_centroid, INTERP_MODE_SMOOTH); case SYSTEM_VALUE_BARYCENTRIC_PERSP_SAMPLE: return nir_load_barycentric(b, nir_intrinsic_load_barycentric_sample, INTERP_MODE_SMOOTH); case SYSTEM_VALUE_BARYCENTRIC_PULL_MODEL: return nir_load_barycentric(b, nir_intrinsic_load_barycentric_model, INTERP_MODE_NONE); case SYSTEM_VALUE_HELPER_INVOCATION: { /* When demote operation is used, reading the HelperInvocation * needs to use Volatile memory access semantics to provide the * correct (dynamic) value. See OpDemoteToHelperInvocation. */ if (nir_intrinsic_access(intrin) & ACCESS_VOLATILE) return nir_is_helper_invocation(b, 1); break; } case SYSTEM_VALUE_MESH_VIEW_INDICES: return nir_load_mesh_view_indices(b, intrin->dest.ssa.num_components, bit_size, column, .base = 0, .range = intrin->dest.ssa.num_components * bit_size / 8); default: break; } nir_intrinsic_op sysval_op = nir_intrinsic_from_system_value(var->data.location); if (glsl_type_is_matrix(var->type)) { assert(nir_intrinsic_infos[sysval_op].index_map[NIR_INTRINSIC_COLUMN] > 0); unsigned num_cols = glsl_get_matrix_columns(var->type); ASSERTED unsigned num_rows = glsl_get_vector_elements(var->type); assert(num_rows == intrin->dest.ssa.num_components); nir_ssa_def *cols[4]; for (unsigned i = 0; i < num_cols; i++) { cols[i] = nir_load_system_value(b, sysval_op, i, intrin->dest.ssa.num_components, intrin->dest.ssa.bit_size); assert(cols[i]->num_components == num_rows); } return nir_select_from_ssa_def_array(b, cols, num_cols, column); } else if (glsl_type_is_array(var->type)) { unsigned num_elems = glsl_get_length(var->type); ASSERTED const struct glsl_type *elem_type = glsl_get_array_element(var->type); assert(glsl_get_components(elem_type) == intrin->dest.ssa.num_components); nir_ssa_def *elems[4]; assert(ARRAY_SIZE(elems) >= num_elems); for (unsigned i = 0; i < num_elems; i++) { elems[i] = nir_load_system_value(b, sysval_op, i, intrin->dest.ssa.num_components, intrin->dest.ssa.bit_size); } return nir_select_from_ssa_def_array(b, elems, num_elems, column); } else { return nir_load_system_value(b, sysval_op, 0, intrin->dest.ssa.num_components, intrin->dest.ssa.bit_size); } } default: return NULL; } } nir_ssa_def * nir_build_lowered_load_helper_invocation(nir_builder *b) { nir_ssa_def *tmp; tmp = nir_ishl(b, nir_imm_int(b, 1), nir_load_sample_id_no_per_sample(b)); tmp = nir_iand(b, nir_load_sample_mask_in(b), tmp); return nir_inot(b, nir_i2b(b, tmp)); } bool nir_lower_system_values(nir_shader *shader) { bool progress = nir_shader_lower_instructions(shader, lower_system_value_filter, lower_system_value_instr, NULL); /* We're going to delete the variables so we need to clean up all those * derefs we left lying around. */ if (progress) nir_remove_dead_derefs(shader); nir_foreach_variable_with_modes_safe(var, shader, nir_var_system_value) exec_node_remove(&var->node); return progress; } static nir_ssa_def * lower_id_to_index_no_umod(nir_builder *b, nir_ssa_def *index, nir_ssa_def *size, unsigned bit_size) { /* We lower ID to Index with the following formula: * * id.z = index / (size.x * size.y) * id.y = (index - (id.z * (size.x * size.y))) / size.x * id.x = index - ((id.z * (size.x * size.y)) + (id.y * size.x)) * * This is more efficient on HW that doesn't have a * modulo division instruction and when the size is either * not compile time known or not a power of two. */ nir_ssa_def *size_x = nir_channel(b, size, 0); nir_ssa_def *size_y = nir_channel(b, size, 1); nir_ssa_def *size_x_y = nir_imul(b, size_x, size_y); nir_ssa_def *id_z = nir_udiv(b, index, size_x_y); nir_ssa_def *z_portion = nir_imul(b, id_z, size_x_y); nir_ssa_def *id_y = nir_udiv(b, nir_isub(b, index, z_portion), size_x); nir_ssa_def *y_portion = nir_imul(b, id_y, size_x); nir_ssa_def *id_x = nir_isub(b, index, nir_iadd(b, z_portion, y_portion)); return nir_u2uN(b, nir_vec3(b, id_x, id_y, id_z), bit_size); } static nir_ssa_def * lower_id_to_index(nir_builder *b, nir_ssa_def *index, nir_ssa_def *size, unsigned bit_size) { /* We lower gl_LocalInvocationID to gl_LocalInvocationIndex based * on this formula: * * id.x = index % size.x; * id.y = (index / size.x) % gl_WorkGroupSize.y; * id.z = (index / (size.x * size.y)) % size.z; * * However, the final % size.z does nothing unless we * accidentally end up with an index that is too * large so it can safely be omitted. * * Because no hardware supports a local workgroup size greater than * about 1K, this calculation can be done in 32-bit and can save some * 64-bit arithmetic. */ nir_ssa_def *size_x = nir_channel(b, size, 0); nir_ssa_def *size_y = nir_channel(b, size, 1); nir_ssa_def *id_x = nir_umod(b, index, size_x); nir_ssa_def *id_y = nir_umod(b, nir_udiv(b, index, size_x), size_y); nir_ssa_def *id_z = nir_udiv(b, index, nir_imul(b, size_x, size_y)); return nir_u2uN(b, nir_vec3(b, id_x, id_y, id_z), bit_size); } static bool lower_compute_system_value_filter(const nir_instr *instr, const void *_state) { return instr->type == nir_instr_type_intrinsic; } static nir_ssa_def * lower_compute_system_value_instr(nir_builder *b, nir_instr *instr, void *_state) { nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); struct lower_sysval_state *state = (struct lower_sysval_state *)_state; const nir_lower_compute_system_values_options *options = state->options; /* All the intrinsics we care about are loads */ if (!nir_intrinsic_infos[intrin->intrinsic].has_dest) return NULL; assert(intrin->dest.is_ssa); const unsigned bit_size = intrin->dest.ssa.bit_size; switch (intrin->intrinsic) { case nir_intrinsic_load_local_invocation_id: /* If lower_cs_local_id_to_index is true, then we replace * local_invocation_id with a formula based on local_invocation_index. */ if (b->shader->options->lower_cs_local_id_to_index || (options && options->lower_cs_local_id_to_index)) { nir_ssa_def *local_index = nir_load_local_invocation_index(b); if (!b->shader->info.workgroup_size_variable) { /* Shortcut for 1 dimensional workgroups: * Use local_invocation_index directly, which is better than * lower_id_to_index + constant folding, because * this way we don't leave behind extra ALU instrs. */ /* size_x = 1, size_y = 1, therefore Z = local index */ if (b->shader->info.workgroup_size[0] == 1 && b->shader->info.workgroup_size[1] == 1) return nir_vec3(b, nir_imm_int(b, 0), nir_imm_int(b, 0), local_index); /* size_x = 1, size_z = 1, therefore Y = local index */ if (b->shader->info.workgroup_size[0] == 1 && b->shader->info.workgroup_size[2] == 1) return nir_vec3(b, nir_imm_int(b, 0), local_index, nir_imm_int(b, 0)); /* size_y = 1, size_z = 1, therefore X = local index */ if (b->shader->info.workgroup_size[1] == 1 && b->shader->info.workgroup_size[2] == 1) return nir_vec3(b, local_index, nir_imm_int(b, 0), nir_imm_int(b, 0)); } nir_ssa_def *local_size = nir_load_workgroup_size(b); return lower_id_to_index(b, local_index, local_size, bit_size); } if (options && options->shuffle_local_ids_for_quad_derivatives && b->shader->info.cs.derivative_group == DERIVATIVE_GROUP_QUADS && _mesa_set_search(state->lower_once_list, instr) == NULL) { nir_ssa_def *ids = nir_load_local_invocation_id(b); _mesa_set_add(state->lower_once_list, ids->parent_instr); nir_ssa_def *x = nir_channel(b, ids, 0); nir_ssa_def *y = nir_channel(b, ids, 1); nir_ssa_def *z = nir_channel(b, ids, 2); unsigned size_x = b->shader->info.workgroup_size[0]; nir_ssa_def *size_x_imm; if (b->shader->info.workgroup_size_variable) size_x_imm = nir_channel(b, nir_load_workgroup_size(b), 0); else size_x_imm = nir_imm_int(b, size_x); /* Remap indices from: * | 0| 1| 2| 3| * | 4| 5| 6| 7| * | 8| 9|10|11| * |12|13|14|15| * to: * | 0| 1| 4| 5| * | 2| 3| 6| 7| * | 8| 9|12|13| * |10|11|14|15| * * That's the layout required by AMD hardware for derivatives to * work. Other hardware may work differently. * * It's a classic tiling pattern that can be implemented by inserting * bit y[0] between bits x[0] and x[1] like this: * * x[0],y[0],x[1],...x[last],y[1],...,y[last] * * If the width is a power of two, use: * i = ((x & 1) | ((y & 1) << 1) | ((x & ~1) << 1)) | ((y & ~1) << logbase2(size_x)) * * If the width is not a power of two or the local size is variable, use: * i = ((x & 1) | ((y & 1) << 1) | ((x & ~1) << 1)) + ((y & ~1) * size_x) * * GL_NV_compute_shader_derivatives requires that the width and height * are a multiple of two, which is also a requirement for the second * expression to work. * * The 2D result is: (x,y) = (i % w, i / w) */ nir_ssa_def *one = nir_imm_int(b, 1); nir_ssa_def *inv_one = nir_imm_int(b, ~1); nir_ssa_def *x_bit0 = nir_iand(b, x, one); nir_ssa_def *y_bit0 = nir_iand(b, y, one); nir_ssa_def *x_bits_1n = nir_iand(b, x, inv_one); nir_ssa_def *y_bits_1n = nir_iand(b, y, inv_one); nir_ssa_def *bits_01 = nir_ior(b, x_bit0, nir_ishl(b, y_bit0, one)); nir_ssa_def *bits_01x = nir_ior(b, bits_01, nir_ishl(b, x_bits_1n, one)); nir_ssa_def *i; if (!b->shader->info.workgroup_size_variable && util_is_power_of_two_nonzero(size_x)) { nir_ssa_def *log2_size_x = nir_imm_int(b, util_logbase2(size_x)); i = nir_ior(b, bits_01x, nir_ishl(b, y_bits_1n, log2_size_x)); } else { i = nir_iadd(b, bits_01x, nir_imul(b, y_bits_1n, size_x_imm)); } /* This should be fast if size_x is an immediate or even a power * of two. */ x = nir_umod(b, i, size_x_imm); y = nir_udiv(b, i, size_x_imm); return nir_vec3(b, x, y, z); } /* If a workgroup size dimension is 1, then the local invocation id must be zero. */ nir_component_mask_t is_zero = 0; is_zero |= b->shader->info.workgroup_size[0] == 1 ? 0x1 : 0x0; is_zero |= b->shader->info.workgroup_size[1] == 1 ? 0x2 : 0x0; is_zero |= b->shader->info.workgroup_size[2] == 1 ? 0x4 : 0x0; if (!b->shader->info.workgroup_size_variable && is_zero) { nir_ssa_scalar defs[3]; for (unsigned i = 0; i < 3; i++) { defs[i] = is_zero & (1 << i) ? nir_get_ssa_scalar(nir_imm_zero(b, 1, 32), 0) : nir_get_ssa_scalar(&intrin->dest.ssa, i); } return nir_vec_scalars(b, defs, 3); } return NULL; case nir_intrinsic_load_local_invocation_index: /* If lower_cs_local_index_to_id is true, then we replace * local_invocation_index with a formula based on local_invocation_id. */ if (b->shader->options->lower_cs_local_index_to_id || (options && options->lower_local_invocation_index)) { /* From the GLSL man page for gl_LocalInvocationIndex: * * "The value of gl_LocalInvocationIndex is equal to * gl_LocalInvocationID.z * gl_WorkGroupSize.x * * gl_WorkGroupSize.y + gl_LocalInvocationID.y * * gl_WorkGroupSize.x + gl_LocalInvocationID.x" */ nir_ssa_def *local_id = nir_load_local_invocation_id(b); nir_ssa_def *local_size = nir_load_workgroup_size(b); nir_ssa_def *size_x = nir_channel(b, local_size, 0); nir_ssa_def *size_y = nir_channel(b, local_size, 1); /* Because no hardware supports a local workgroup size greater than * about 1K, this calculation can be done in 32-bit and can save some * 64-bit arithmetic. */ nir_ssa_def *index; index = nir_imul(b, nir_channel(b, local_id, 2), nir_imul(b, size_x, size_y)); index = nir_iadd(b, index, nir_imul(b, nir_channel(b, local_id, 1), size_x)); index = nir_iadd(b, index, nir_channel(b, local_id, 0)); return nir_u2uN(b, index, bit_size); } else { return NULL; } case nir_intrinsic_load_workgroup_size: if (b->shader->info.workgroup_size_variable) { /* If the local work group size is variable it can't be lowered at * this point. We do, however, have to make sure that the intrinsic * is only 32-bit. */ return NULL; } else { /* using a 32 bit constant is safe here as no device/driver needs more * than 32 bits for the local size */ nir_const_value workgroup_size_const[3]; memset(workgroup_size_const, 0, sizeof(workgroup_size_const)); workgroup_size_const[0].u32 = b->shader->info.workgroup_size[0]; workgroup_size_const[1].u32 = b->shader->info.workgroup_size[1]; workgroup_size_const[2].u32 = b->shader->info.workgroup_size[2]; return nir_u2uN(b, nir_build_imm(b, 3, 32, workgroup_size_const), bit_size); } case nir_intrinsic_load_global_invocation_id_zero_base: { if ((options && options->has_base_workgroup_id) || !b->shader->options->has_cs_global_id) { nir_ssa_def *group_size = nir_load_workgroup_size(b); nir_ssa_def *group_id = nir_load_workgroup_id(b, bit_size); nir_ssa_def *local_id = nir_load_local_invocation_id(b); return nir_iadd(b, nir_imul(b, group_id, nir_u2uN(b, group_size, bit_size)), nir_u2uN(b, local_id, bit_size)); } else { return NULL; } } case nir_intrinsic_load_global_invocation_id: { if (options && options->has_base_global_invocation_id) return nir_iadd(b, nir_load_global_invocation_id_zero_base(b, bit_size), nir_load_base_global_invocation_id(b, bit_size)); else if ((options && options->has_base_workgroup_id) || !b->shader->options->has_cs_global_id) return nir_load_global_invocation_id_zero_base(b, bit_size); else return NULL; } case nir_intrinsic_load_global_invocation_index: { /* OpenCL's global_linear_id explicitly removes the global offset before computing this */ assert(b->shader->info.stage == MESA_SHADER_KERNEL); nir_ssa_def *global_base_id = nir_load_base_global_invocation_id(b, bit_size); nir_ssa_def *global_id = nir_isub(b, nir_load_global_invocation_id(b, bit_size), global_base_id); nir_ssa_def *global_size = build_global_group_size(b, bit_size); /* index = id.x + ((id.y + (id.z * size.y)) * size.x) */ nir_ssa_def *index; index = nir_imul(b, nir_channel(b, global_id, 2), nir_channel(b, global_size, 1)); index = nir_iadd(b, nir_channel(b, global_id, 1), index); index = nir_imul(b, nir_channel(b, global_size, 0), index); index = nir_iadd(b, nir_channel(b, global_id, 0), index); return index; } case nir_intrinsic_load_workgroup_id: { if (options && options->has_base_workgroup_id) return nir_iadd(b, nir_u2uN(b, nir_load_workgroup_id_zero_base(b), bit_size), nir_load_base_workgroup_id(b, bit_size)); else if (options && options->lower_workgroup_id_to_index) return lower_id_to_index_no_umod(b, nir_load_workgroup_index(b), nir_load_num_workgroups(b, bit_size), bit_size); return NULL; } default: return NULL; } } bool nir_lower_compute_system_values(nir_shader *shader, const nir_lower_compute_system_values_options *options) { if (!gl_shader_stage_uses_workgroup(shader->info.stage)) return false; struct lower_sysval_state state; state.options = options; state.lower_once_list = _mesa_pointer_set_create(NULL); bool progress = nir_shader_lower_instructions(shader, lower_compute_system_value_filter, lower_compute_system_value_instr, (void*)&state); ralloc_free(state.lower_once_list); /* Update this so as not to lower it again. */ if (options && options->shuffle_local_ids_for_quad_derivatives && shader->info.cs.derivative_group == DERIVATIVE_GROUP_QUADS) shader->info.cs.derivative_group = DERIVATIVE_GROUP_LINEAR; return progress; }