/* * Copyright © Microsoft 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 "nir.h" #include "nir_serialize.h" #include "glsl_types.h" #include "nir_types.h" #include "clc_compiler.h" #include "clc_helpers.h" #include "clc_nir.h" #include "../compiler/dxil_nir.h" #include "../compiler/dxil_nir_lower_int_samplers.h" #include "../compiler/nir_to_dxil.h" #include "util/u_debug.h" #include #include "spirv/nir_spirv.h" #include "nir_builder.h" #include "nir_builtin_builder.h" #include "git_sha1.h" struct clc_image_lower_context { struct clc_dxil_metadata *metadata; unsigned *num_srvs; unsigned *num_uavs; nir_deref_instr *deref; unsigned num_buf_ids; int metadata_index; }; static int lower_image_deref_impl(nir_builder *b, struct clc_image_lower_context *context, const struct glsl_type *new_var_type, unsigned *num_bindings) { nir_variable *in_var = nir_deref_instr_get_variable(context->deref); nir_variable *uniform = nir_variable_create(b->shader, nir_var_uniform, new_var_type, NULL); uniform->data.access = in_var->data.access; uniform->data.binding = in_var->data.binding; if (context->num_buf_ids > 0) { // Need to assign a new binding context->metadata->args[context->metadata_index]. image.buf_ids[context->num_buf_ids] = uniform->data.binding = (*num_bindings)++; } context->num_buf_ids++; return uniform->data.binding; } static int lower_read_only_image_deref(nir_builder *b, struct clc_image_lower_context *context, nir_alu_type image_type) { nir_variable *in_var = nir_deref_instr_get_variable(context->deref); // Non-writeable images should be converted to samplers, // since they may have texture operations done on them const struct glsl_type *new_var_type = glsl_sampler_type(glsl_get_sampler_dim(in_var->type), false, glsl_sampler_type_is_array(in_var->type), nir_get_glsl_base_type_for_nir_type(image_type | 32)); return lower_image_deref_impl(b, context, new_var_type, context->num_srvs); } static int lower_read_write_image_deref(nir_builder *b, struct clc_image_lower_context *context, nir_alu_type image_type) { nir_variable *in_var = nir_deref_instr_get_variable(context->deref); const struct glsl_type *new_var_type = glsl_image_type(glsl_get_sampler_dim(in_var->type), glsl_sampler_type_is_array(in_var->type), nir_get_glsl_base_type_for_nir_type(image_type | 32)); return lower_image_deref_impl(b, context, new_var_type, context->num_uavs); } static void clc_lower_input_image_deref(nir_builder *b, struct clc_image_lower_context *context) { // The input variable here isn't actually an image, it's just the // image format data. // // For every use of an image in a different way, we'll add an // appropriate uniform to match it. That can result in up to // 3 uniforms (float4, int4, uint4) for each image. Only one of these // formats will actually produce correct data, but a single kernel // could use runtime conditionals to potentially access any of them. // // If the image is used in a query that doesn't have a corresponding // DXIL intrinsic (CL image channel order or channel format), then // we'll add a kernel input for that data that'll be lowered by the // explicit IO pass later on. // // After all that, we can remove the image input variable and deref. enum image_uniform_type { FLOAT4, INT4, UINT4, IMAGE_UNIFORM_TYPE_COUNT }; int image_bindings[IMAGE_UNIFORM_TYPE_COUNT] = {-1, -1, -1}; nir_ssa_def *format_deref_dest = NULL, *order_deref_dest = NULL; nir_variable *in_var = nir_deref_instr_get_variable(context->deref); enum gl_access_qualifier access = in_var->data.access; context->metadata_index = 0; while (context->metadata->args[context->metadata_index].image.buf_ids[0] != in_var->data.binding) context->metadata_index++; context->num_buf_ids = 0; /* Do this in 2 passes: * 1. When encountering a strongly-typed access (load/store), replace the deref * with one that references an appropriately typed variable. When encountering * an untyped access (size query), if we have a strongly-typed variable already, * replace the deref to point to it. * 2. If there's any references left, they should all be untyped. If we found * a strongly-typed access later in the 1st pass, then just replace the reference. * If we didn't, e.g. the resource is only used for a size query, then pick an * arbitrary type for it. */ for (int pass = 0; pass < 2; ++pass) { nir_foreach_use_safe(src, &context->deref->dest.ssa) { enum image_uniform_type type; if (src->parent_instr->type == nir_instr_type_intrinsic) { nir_intrinsic_instr *intrinsic = nir_instr_as_intrinsic(src->parent_instr); enum nir_alu_type dest_type; b->cursor = nir_before_instr(&intrinsic->instr); switch (intrinsic->intrinsic) { case nir_intrinsic_image_deref_load: case nir_intrinsic_image_deref_store: { dest_type = intrinsic->intrinsic == nir_intrinsic_image_deref_load ? nir_intrinsic_dest_type(intrinsic) : nir_intrinsic_src_type(intrinsic); switch (nir_alu_type_get_base_type(dest_type)) { case nir_type_float: type = FLOAT4; break; case nir_type_int: type = INT4; break; case nir_type_uint: type = UINT4; break; default: unreachable("Unsupported image type for load."); } int image_binding = image_bindings[type]; if (image_binding < 0) { image_binding = image_bindings[type] = lower_read_write_image_deref(b, context, dest_type); } assert((in_var->data.access & ACCESS_NON_WRITEABLE) == 0); nir_rewrite_image_intrinsic(intrinsic, nir_imm_int(b, image_binding), false); break; } case nir_intrinsic_image_deref_size: { int image_binding = -1; for (unsigned i = 0; i < IMAGE_UNIFORM_TYPE_COUNT; ++i) { if (image_bindings[i] >= 0) { image_binding = image_bindings[i]; break; } } if (image_binding < 0) { // Skip for now and come back to it if (pass == 0) break; type = FLOAT4; image_binding = image_bindings[type] = lower_read_write_image_deref(b, context, nir_type_float32); } assert((in_var->data.access & ACCESS_NON_WRITEABLE) == 0); nir_rewrite_image_intrinsic(intrinsic, nir_imm_int(b, image_binding), false); break; } case nir_intrinsic_image_deref_format: case nir_intrinsic_image_deref_order: { nir_ssa_def **cached_deref = intrinsic->intrinsic == nir_intrinsic_image_deref_format ? &format_deref_dest : &order_deref_dest; if (!*cached_deref) { nir_variable *new_input = nir_variable_create(b->shader, nir_var_uniform, glsl_uint_type(), NULL); new_input->data.driver_location = in_var->data.driver_location; if (intrinsic->intrinsic == nir_intrinsic_image_deref_format) { /* Match cl_image_format { image_channel_order, image_channel_data_type }; */ new_input->data.driver_location += glsl_get_cl_size(new_input->type); } b->cursor = nir_after_instr(&context->deref->instr); *cached_deref = nir_load_var(b, new_input); } /* No actual intrinsic needed here, just reference the loaded variable */ nir_ssa_def_rewrite_uses(&intrinsic->dest.ssa, *cached_deref); nir_instr_remove(&intrinsic->instr); break; } default: unreachable("Unsupported image intrinsic"); } } else if (src->parent_instr->type == nir_instr_type_tex) { assert(in_var->data.access & ACCESS_NON_WRITEABLE); nir_tex_instr *tex = nir_instr_as_tex(src->parent_instr); switch (nir_alu_type_get_base_type(tex->dest_type)) { case nir_type_float: type = FLOAT4; break; case nir_type_int: type = INT4; break; case nir_type_uint: type = UINT4; break; default: unreachable("Unsupported image format for sample."); } int image_binding = image_bindings[type]; if (image_binding < 0) { image_binding = image_bindings[type] = lower_read_only_image_deref(b, context, tex->dest_type); } nir_tex_instr_remove_src(tex, nir_tex_instr_src_index(tex, nir_tex_src_texture_deref)); tex->texture_index = image_binding; } } } context->metadata->args[context->metadata_index].image.num_buf_ids = context->num_buf_ids; nir_instr_remove(&context->deref->instr); exec_node_remove(&in_var->node); } static void clc_lower_images(nir_shader *nir, struct clc_image_lower_context *context) { nir_foreach_function(func, nir) { if (!func->is_entrypoint) continue; assert(func->impl); nir_builder b; nir_builder_init(&b, func->impl); nir_foreach_block(block, func->impl) { nir_foreach_instr_safe(instr, block) { if (instr->type == nir_instr_type_deref) { context->deref = nir_instr_as_deref(instr); if (glsl_type_is_image(context->deref->type)) { assert(context->deref->deref_type == nir_deref_type_var); clc_lower_input_image_deref(&b, context); } } } } } } static void clc_lower_64bit_semantics(nir_shader *nir) { nir_foreach_function(func, nir) { nir_builder b; nir_builder_init(&b, func->impl); nir_foreach_block(block, func->impl) { nir_foreach_instr_safe(instr, block) { if (instr->type == nir_instr_type_intrinsic) { nir_intrinsic_instr *intrinsic = nir_instr_as_intrinsic(instr); switch (intrinsic->intrinsic) { case nir_intrinsic_load_global_invocation_id: case nir_intrinsic_load_global_invocation_id_zero_base: case nir_intrinsic_load_base_global_invocation_id: case nir_intrinsic_load_local_invocation_id: case nir_intrinsic_load_workgroup_id: case nir_intrinsic_load_workgroup_id_zero_base: case nir_intrinsic_load_base_workgroup_id: case nir_intrinsic_load_num_workgroups: break; default: continue; } if (nir_instr_ssa_def(instr)->bit_size != 64) continue; intrinsic->dest.ssa.bit_size = 32; b.cursor = nir_after_instr(instr); nir_ssa_def *i64 = nir_u2u64(&b, &intrinsic->dest.ssa); nir_ssa_def_rewrite_uses_after( &intrinsic->dest.ssa, i64, i64->parent_instr); } } } } } static void clc_lower_nonnormalized_samplers(nir_shader *nir, const dxil_wrap_sampler_state *states) { nir_foreach_function(func, nir) { if (!func->is_entrypoint) continue; assert(func->impl); nir_builder b; nir_builder_init(&b, func->impl); nir_foreach_block(block, func->impl) { nir_foreach_instr_safe(instr, block) { if (instr->type != nir_instr_type_tex) continue; nir_tex_instr *tex = nir_instr_as_tex(instr); int sampler_src_idx = nir_tex_instr_src_index(tex, nir_tex_src_sampler_deref); if (sampler_src_idx == -1) continue; nir_src *sampler_src = &tex->src[sampler_src_idx].src; assert(sampler_src->is_ssa && sampler_src->ssa->parent_instr->type == nir_instr_type_deref); nir_variable *sampler = nir_deref_instr_get_variable( nir_instr_as_deref(sampler_src->ssa->parent_instr)); // If the sampler returns ints, we'll handle this in the int lowering pass if (nir_alu_type_get_base_type(tex->dest_type) != nir_type_float) continue; // If sampler uses normalized coords, nothing to do if (!states[sampler->data.binding].is_nonnormalized_coords) continue; b.cursor = nir_before_instr(&tex->instr); int coords_idx = nir_tex_instr_src_index(tex, nir_tex_src_coord); assert(coords_idx != -1); nir_ssa_def *coords = nir_ssa_for_src(&b, tex->src[coords_idx].src, tex->coord_components); nir_ssa_def *txs = nir_i2f32(&b, nir_get_texture_size(&b, tex)); // Normalize coords for tex nir_ssa_def *scale = nir_frcp(&b, txs); nir_ssa_def *comps[4]; for (unsigned i = 0; i < coords->num_components; ++i) { comps[i] = nir_channel(&b, coords, i); if (tex->is_array && i == coords->num_components - 1) { // Don't scale the array index, but do clamp it comps[i] = nir_fround_even(&b, comps[i]); comps[i] = nir_fmax(&b, comps[i], nir_imm_float(&b, 0.0f)); comps[i] = nir_fmin(&b, comps[i], nir_fsub(&b, nir_channel(&b, txs, i), nir_imm_float(&b, 1.0f))); break; } // The CTS is pretty clear that this value has to be floored for nearest sampling // but must not be for linear sampling. if (!states[sampler->data.binding].is_linear_filtering) comps[i] = nir_fadd_imm(&b, nir_ffloor(&b, comps[i]), 0.5f); comps[i] = nir_fmul(&b, comps[i], nir_channel(&b, scale, i)); } nir_ssa_def *normalized_coords = nir_vec(&b, comps, coords->num_components); nir_instr_rewrite_src(&tex->instr, &tex->src[coords_idx].src, nir_src_for_ssa(normalized_coords)); } } } } static nir_variable * add_kernel_inputs_var(struct clc_dxil_object *dxil, nir_shader *nir, unsigned *cbv_id) { if (!dxil->kernel->num_args) return NULL; struct clc_dxil_metadata *metadata = &dxil->metadata; unsigned size = 0; nir_foreach_variable_with_modes(var, nir, nir_var_uniform) size = MAX2(size, var->data.driver_location + glsl_get_cl_size(var->type)); size = align(size, 4); const struct glsl_type *array_type = glsl_array_type(glsl_uint_type(), size / 4, 4); const struct glsl_struct_field field = { array_type, "arr" }; nir_variable *var = nir_variable_create(nir, nir_var_mem_ubo, glsl_struct_type(&field, 1, "kernel_inputs", false), "kernel_inputs"); var->data.binding = (*cbv_id)++; var->data.how_declared = nir_var_hidden; return var; } static nir_variable * add_work_properties_var(struct clc_dxil_object *dxil, struct nir_shader *nir, unsigned *cbv_id) { struct clc_dxil_metadata *metadata = &dxil->metadata; const struct glsl_type *array_type = glsl_array_type(glsl_uint_type(), sizeof(struct clc_work_properties_data) / sizeof(unsigned), sizeof(unsigned)); const struct glsl_struct_field field = { array_type, "arr" }; nir_variable *var = nir_variable_create(nir, nir_var_mem_ubo, glsl_struct_type(&field, 1, "kernel_work_properties", false), "kernel_work_properies"); var->data.binding = (*cbv_id)++; var->data.how_declared = nir_var_hidden; return var; } static void clc_lower_constant_to_ssbo(nir_shader *nir, const struct clc_kernel_info *kerninfo, unsigned *uav_id) { /* Update UBO vars and assign them a binding. */ nir_foreach_variable_with_modes(var, nir, nir_var_mem_constant) { var->data.mode = nir_var_mem_ssbo; var->data.binding = (*uav_id)++; } /* And finally patch all the derefs referincing the constant * variables/pointers. */ nir_foreach_function(func, nir) { if (!func->is_entrypoint) continue; assert(func->impl); nir_builder b; nir_builder_init(&b, func->impl); nir_foreach_block(block, func->impl) { nir_foreach_instr(instr, block) { if (instr->type != nir_instr_type_deref) continue; nir_deref_instr *deref = nir_instr_as_deref(instr); if (deref->modes != nir_var_mem_constant) continue; deref->modes = nir_var_mem_ssbo; } } } } static void clc_lower_global_to_ssbo(nir_shader *nir) { nir_foreach_function(func, nir) { if (!func->is_entrypoint) continue; assert(func->impl); nir_foreach_block(block, func->impl) { nir_foreach_instr(instr, block) { if (instr->type != nir_instr_type_deref) continue; nir_deref_instr *deref = nir_instr_as_deref(instr); if (deref->modes != nir_var_mem_global) continue; deref->modes = nir_var_mem_ssbo; } } } } static void copy_const_initializer(const nir_constant *constant, const struct glsl_type *type, uint8_t *data) { unsigned size = glsl_get_cl_size(type); if (glsl_type_is_array(type)) { const struct glsl_type *elm_type = glsl_get_array_element(type); unsigned step_size = glsl_get_explicit_stride(type); for (unsigned i = 0; i < constant->num_elements; i++) { copy_const_initializer(constant->elements[i], elm_type, data + (i * step_size)); } } else if (glsl_type_is_struct(type)) { for (unsigned i = 0; i < constant->num_elements; i++) { const struct glsl_type *elm_type = glsl_get_struct_field(type, i); int offset = glsl_get_struct_field_offset(type, i); copy_const_initializer(constant->elements[i], elm_type, data + offset); } } else { assert(glsl_type_is_vector_or_scalar(type)); for (unsigned i = 0; i < glsl_get_components(type); i++) { switch (glsl_get_bit_size(type)) { case 64: *((uint64_t *)data) = constant->values[i].u64; break; case 32: *((uint32_t *)data) = constant->values[i].u32; break; case 16: *((uint16_t *)data) = constant->values[i].u16; break; case 8: *((uint8_t *)data) = constant->values[i].u8; break; default: unreachable("Invalid base type"); } data += glsl_get_bit_size(type) / 8; } } } static const struct glsl_type * get_cast_type(unsigned bit_size) { switch (bit_size) { case 64: return glsl_int64_t_type(); case 32: return glsl_int_type(); case 16: return glsl_int16_t_type(); case 8: return glsl_int8_t_type(); } unreachable("Invalid bit_size"); } static void split_unaligned_load(nir_builder *b, nir_intrinsic_instr *intrin, unsigned alignment) { enum gl_access_qualifier access = nir_intrinsic_access(intrin); nir_ssa_def *srcs[NIR_MAX_VEC_COMPONENTS * NIR_MAX_VEC_COMPONENTS * sizeof(int64_t) / 8]; unsigned comp_size = intrin->dest.ssa.bit_size / 8; unsigned num_comps = intrin->dest.ssa.num_components; b->cursor = nir_before_instr(&intrin->instr); nir_deref_instr *ptr = nir_src_as_deref(intrin->src[0]); const struct glsl_type *cast_type = get_cast_type(alignment * 8); nir_deref_instr *cast = nir_build_deref_cast(b, &ptr->dest.ssa, ptr->modes, cast_type, alignment); unsigned num_loads = DIV_ROUND_UP(comp_size * num_comps, alignment); for (unsigned i = 0; i < num_loads; ++i) { nir_deref_instr *elem = nir_build_deref_ptr_as_array(b, cast, nir_imm_intN_t(b, i, cast->dest.ssa.bit_size)); srcs[i] = nir_load_deref_with_access(b, elem, access); } nir_ssa_def *new_dest = nir_extract_bits(b, srcs, num_loads, 0, num_comps, intrin->dest.ssa.bit_size); nir_ssa_def_rewrite_uses(&intrin->dest.ssa, new_dest); nir_instr_remove(&intrin->instr); } static void split_unaligned_store(nir_builder *b, nir_intrinsic_instr *intrin, unsigned alignment) { enum gl_access_qualifier access = nir_intrinsic_access(intrin); assert(intrin->src[1].is_ssa); nir_ssa_def *value = intrin->src[1].ssa; unsigned comp_size = value->bit_size / 8; unsigned num_comps = value->num_components; b->cursor = nir_before_instr(&intrin->instr); nir_deref_instr *ptr = nir_src_as_deref(intrin->src[0]); const struct glsl_type *cast_type = get_cast_type(alignment * 8); nir_deref_instr *cast = nir_build_deref_cast(b, &ptr->dest.ssa, ptr->modes, cast_type, alignment); unsigned num_stores = DIV_ROUND_UP(comp_size * num_comps, alignment); for (unsigned i = 0; i < num_stores; ++i) { nir_ssa_def *substore_val = nir_extract_bits(b, &value, 1, i * alignment * 8, 1, alignment * 8); nir_deref_instr *elem = nir_build_deref_ptr_as_array(b, cast, nir_imm_intN_t(b, i, cast->dest.ssa.bit_size)); nir_store_deref_with_access(b, elem, substore_val, ~0, access); } nir_instr_remove(&intrin->instr); } static bool split_unaligned_loads_stores(nir_shader *shader) { bool progress = false; nir_foreach_function(function, shader) { if (!function->impl) continue; nir_builder b; nir_builder_init(&b, function->impl); nir_foreach_block(block, function->impl) { nir_foreach_instr_safe(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_deref && intrin->intrinsic != nir_intrinsic_store_deref) continue; nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]); unsigned align_mul = 0, align_offset = 0; nir_get_explicit_deref_align(deref, true, &align_mul, &align_offset); unsigned alignment = align_offset ? 1 << (ffs(align_offset) - 1) : align_mul; /* We can load anything at 4-byte alignment, except for * UBOs (AKA CBs where the granularity is 16 bytes). */ if (alignment >= (deref->modes == nir_var_mem_ubo ? 16 : 4)) continue; nir_ssa_def *val; if (intrin->intrinsic == nir_intrinsic_load_deref) { assert(intrin->dest.is_ssa); val = &intrin->dest.ssa; } else { assert(intrin->src[1].is_ssa); val = intrin->src[1].ssa; } unsigned natural_alignment = val->bit_size / 8 * (val->num_components == 3 ? 4 : val->num_components); if (alignment >= natural_alignment) continue; if (intrin->intrinsic == nir_intrinsic_load_deref) split_unaligned_load(&b, intrin, alignment); else split_unaligned_store(&b, intrin, alignment); progress = true; } } } return progress; } static enum pipe_tex_wrap wrap_from_cl_addressing(unsigned addressing_mode) { switch (addressing_mode) { default: case SAMPLER_ADDRESSING_MODE_NONE: case SAMPLER_ADDRESSING_MODE_CLAMP: // Since OpenCL's only border color is 0's and D3D specs out-of-bounds loads to return 0, don't apply any wrap mode return (enum pipe_tex_wrap)-1; case SAMPLER_ADDRESSING_MODE_CLAMP_TO_EDGE: return PIPE_TEX_WRAP_CLAMP_TO_EDGE; case SAMPLER_ADDRESSING_MODE_REPEAT: return PIPE_TEX_WRAP_REPEAT; case SAMPLER_ADDRESSING_MODE_REPEAT_MIRRORED: return PIPE_TEX_WRAP_MIRROR_REPEAT; } } static bool shader_has_double(nir_shader *nir) { bool progress = false; foreach_list_typed(nir_function, func, node, &nir->functions) { if (!func->is_entrypoint) continue; assert(func->impl); nir_foreach_block(block, func->impl) { nir_foreach_instr_safe(instr, block) { if (instr->type != nir_instr_type_alu) continue; nir_alu_instr *alu = nir_instr_as_alu(instr); const nir_op_info *info = &nir_op_infos[alu->op]; if (info->output_type & nir_type_float && nir_dest_bit_size(alu->dest.dest) == 64) return true; } } } return false; } static bool scale_fdiv(nir_shader *nir) { bool progress = false; nir_foreach_function(func, nir) { if (!func->impl) continue; nir_builder b; nir_builder_init(&b, func->impl); nir_foreach_block(block, func->impl) { nir_foreach_instr(instr, block) { if (instr->type != nir_instr_type_alu) continue; nir_alu_instr *alu = nir_instr_as_alu(instr); if (alu->op != nir_op_fdiv || alu->src[0].src.ssa->bit_size != 32) continue; b.cursor = nir_before_instr(instr); nir_ssa_def *fabs = nir_fabs(&b, alu->src[1].src.ssa); nir_ssa_def *big = nir_flt(&b, nir_imm_int(&b, 0x7e800000), fabs); nir_ssa_def *small = nir_flt(&b, fabs, nir_imm_int(&b, 0x00800000)); nir_ssa_def *scaled_down_a = nir_fmul_imm(&b, alu->src[0].src.ssa, 0.25); nir_ssa_def *scaled_down_b = nir_fmul_imm(&b, alu->src[1].src.ssa, 0.25); nir_ssa_def *scaled_up_a = nir_fmul_imm(&b, alu->src[0].src.ssa, 16777216.0); nir_ssa_def *scaled_up_b = nir_fmul_imm(&b, alu->src[1].src.ssa, 16777216.0); nir_ssa_def *final_a = nir_bcsel(&b, big, scaled_down_a, (nir_bcsel(&b, small, scaled_up_a, alu->src[0].src.ssa))); nir_ssa_def *final_b = nir_bcsel(&b, big, scaled_down_b, (nir_bcsel(&b, small, scaled_up_b, alu->src[1].src.ssa))); nir_instr_rewrite_src(instr, &alu->src[0].src, nir_src_for_ssa(final_a)); nir_instr_rewrite_src(instr, &alu->src[1].src, nir_src_for_ssa(final_b)); progress = true; } } } return progress; } struct clc_libclc * clc_libclc_new_dxil(const struct clc_logger *logger, const struct clc_libclc_dxil_options *options) { struct clc_libclc_options clc_options = { .optimize = options->optimize, .nir_options = dxil_get_nir_compiler_options(), }; return clc_libclc_new(logger, &clc_options); } bool clc_spirv_to_dxil(struct clc_libclc *lib, const struct clc_binary *linked_spirv, const struct clc_parsed_spirv *parsed_data, const char *entrypoint, const struct clc_runtime_kernel_conf *conf, const struct clc_spirv_specialization_consts *consts, const struct clc_logger *logger, struct clc_dxil_object *out_dxil) { struct nir_shader *nir; for (unsigned i = 0; i < parsed_data->num_kernels; i++) { if (!strcmp(parsed_data->kernels[i].name, entrypoint)) { out_dxil->kernel = &parsed_data->kernels[i]; break; } } if (!out_dxil->kernel) { clc_error(logger, "no '%s' kernel found", entrypoint); return false; } const struct spirv_to_nir_options spirv_options = { .environment = NIR_SPIRV_OPENCL, .clc_shader = clc_libclc_get_clc_shader(lib), .constant_addr_format = nir_address_format_32bit_index_offset_pack64, .global_addr_format = nir_address_format_32bit_index_offset_pack64, .shared_addr_format = nir_address_format_32bit_offset_as_64bit, .temp_addr_format = nir_address_format_32bit_offset_as_64bit, .float_controls_execution_mode = FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32, .caps = { .address = true, .float64 = true, .int8 = true, .int16 = true, .int64 = true, .kernel = true, .kernel_image = true, .kernel_image_read_write = true, .literal_sampler = true, .printf = true, }, }; nir_shader_compiler_options nir_options = *dxil_get_nir_compiler_options(); if (conf && conf->lower_bit_size & 64) { nir_options.lower_pack_64_2x32_split = false; nir_options.lower_unpack_64_2x32_split = false; nir_options.lower_int64_options = ~0; } if (conf && conf->lower_bit_size & 16) nir_options.support_16bit_alu = true; glsl_type_singleton_init_or_ref(); nir = spirv_to_nir(linked_spirv->data, linked_spirv->size / 4, consts ? (struct nir_spirv_specialization *)consts->specializations : NULL, consts ? consts->num_specializations : 0, MESA_SHADER_KERNEL, entrypoint, &spirv_options, &nir_options); if (!nir) { clc_error(logger, "spirv_to_nir() failed"); goto err_free_dxil; } nir->info.workgroup_size_variable = true; NIR_PASS_V(nir, nir_lower_goto_ifs); NIR_PASS_V(nir, nir_opt_dead_cf); struct clc_dxil_metadata *metadata = &out_dxil->metadata; metadata->args = calloc(out_dxil->kernel->num_args, sizeof(*metadata->args)); if (!metadata->args) { clc_error(logger, "failed to allocate arg positions"); goto err_free_dxil; } { bool progress; do { progress = false; NIR_PASS(progress, nir, nir_copy_prop); NIR_PASS(progress, nir, nir_opt_copy_prop_vars); NIR_PASS(progress, nir, nir_opt_deref); NIR_PASS(progress, nir, nir_opt_dce); NIR_PASS(progress, nir, nir_opt_undef); NIR_PASS(progress, nir, nir_opt_constant_folding); NIR_PASS(progress, nir, nir_opt_cse); NIR_PASS(progress, nir, nir_lower_vars_to_ssa); NIR_PASS(progress, nir, nir_opt_algebraic); } while (progress); } // Inline all functions first. // according to the comment on nir_inline_functions NIR_PASS_V(nir, nir_lower_variable_initializers, nir_var_function_temp); NIR_PASS_V(nir, nir_lower_returns); NIR_PASS_V(nir, nir_lower_libclc, clc_libclc_get_clc_shader(lib)); NIR_PASS_V(nir, nir_inline_functions); // Pick off the single entrypoint that we want. foreach_list_typed_safe(nir_function, func, node, &nir->functions) { if (!func->is_entrypoint) exec_node_remove(&func->node); } assert(exec_list_length(&nir->functions) == 1); { bool progress; do { progress = false; NIR_PASS(progress, nir, nir_copy_prop); NIR_PASS(progress, nir, nir_opt_copy_prop_vars); NIR_PASS(progress, nir, nir_opt_deref); NIR_PASS(progress, nir, nir_opt_dce); NIR_PASS(progress, nir, nir_opt_undef); NIR_PASS(progress, nir, nir_opt_constant_folding); NIR_PASS(progress, nir, nir_opt_cse); NIR_PASS(progress, nir, nir_split_var_copies); NIR_PASS(progress, nir, nir_lower_var_copies); NIR_PASS(progress, nir, nir_lower_vars_to_ssa); NIR_PASS(progress, nir, nir_opt_algebraic); NIR_PASS(progress, nir, nir_opt_if, true); NIR_PASS(progress, nir, nir_opt_dead_cf); NIR_PASS(progress, nir, nir_opt_remove_phis); NIR_PASS(progress, nir, nir_opt_peephole_select, 8, true, true); NIR_PASS(progress, nir, nir_lower_vec3_to_vec4, nir_var_mem_generic | nir_var_uniform); } while (progress); } NIR_PASS_V(nir, scale_fdiv); dxil_wrap_sampler_state int_sampler_states[PIPE_MAX_SHADER_SAMPLER_VIEWS] = { {{0}} }; unsigned sampler_id = 0; struct exec_list inline_samplers_list; exec_list_make_empty(&inline_samplers_list); // Move inline samplers to the end of the uniforms list nir_foreach_variable_with_modes_safe(var, nir, nir_var_uniform) { if (glsl_type_is_sampler(var->type) && var->data.sampler.is_inline_sampler) { exec_node_remove(&var->node); exec_list_push_tail(&inline_samplers_list, &var->node); } } exec_node_insert_list_after(exec_list_get_tail(&nir->variables), &inline_samplers_list); NIR_PASS_V(nir, nir_lower_variable_initializers, ~(nir_var_function_temp | nir_var_shader_temp)); // Lower memcpy NIR_PASS_V(nir, dxil_nir_lower_memcpy_deref); // Ensure the printf struct has explicit types, but we'll throw away the scratch size, because we haven't // necessarily removed all temp variables (e.g. the printf struct itself) at this point, so we'll rerun this later assert(nir->scratch_size == 0); NIR_PASS_V(nir, nir_lower_vars_to_explicit_types, nir_var_function_temp, glsl_get_cl_type_size_align); nir_lower_printf_options printf_options = { .treat_doubles_as_floats = true, .max_buffer_size = 1024 * 1024 }; NIR_PASS_V(nir, nir_lower_printf, &printf_options); metadata->printf.info_count = nir->printf_info_count; metadata->printf.infos = calloc(nir->printf_info_count, sizeof(struct clc_printf_info)); for (unsigned i = 0; i < nir->printf_info_count; i++) { metadata->printf.infos[i].str = malloc(nir->printf_info[i].string_size); memcpy(metadata->printf.infos[i].str, nir->printf_info[i].strings, nir->printf_info[i].string_size); metadata->printf.infos[i].num_args = nir->printf_info[i].num_args; metadata->printf.infos[i].arg_sizes = malloc(nir->printf_info[i].num_args * sizeof(unsigned)); memcpy(metadata->printf.infos[i].arg_sizes, nir->printf_info[i].arg_sizes, nir->printf_info[i].num_args * sizeof(unsigned)); } // copy propagate to prepare for lower_explicit_io NIR_PASS_V(nir, nir_split_var_copies); NIR_PASS_V(nir, nir_opt_copy_prop_vars); NIR_PASS_V(nir, nir_lower_var_copies); NIR_PASS_V(nir, nir_lower_vars_to_ssa); NIR_PASS_V(nir, nir_lower_alu); NIR_PASS_V(nir, nir_opt_dce); NIR_PASS_V(nir, nir_opt_deref); // For uniforms (kernel inputs), run this before adjusting variable list via image/sampler lowering NIR_PASS_V(nir, nir_lower_vars_to_explicit_types, nir_var_uniform, glsl_get_cl_type_size_align); // Calculate input offsets/metadata. unsigned uav_id = 0; nir_foreach_variable_with_modes(var, nir, nir_var_uniform) { int i = var->data.location; if (i < 0) continue; unsigned size = glsl_get_cl_size(var->type); metadata->args[i].offset = var->data.driver_location; metadata->args[i].size = size; metadata->kernel_inputs_buf_size = MAX2(metadata->kernel_inputs_buf_size, var->data.driver_location + size); if ((out_dxil->kernel->args[i].address_qualifier == CLC_KERNEL_ARG_ADDRESS_GLOBAL || out_dxil->kernel->args[i].address_qualifier == CLC_KERNEL_ARG_ADDRESS_CONSTANT) && // Ignore images during this pass - global memory buffers need to have contiguous bindings !glsl_type_is_image(var->type)) { metadata->args[i].globconstptr.buf_id = uav_id++; } else if (glsl_type_is_sampler(var->type)) { unsigned address_mode = conf ? conf->args[i].sampler.addressing_mode : 0u; int_sampler_states[sampler_id].wrap[0] = int_sampler_states[sampler_id].wrap[1] = int_sampler_states[sampler_id].wrap[2] = wrap_from_cl_addressing(address_mode); int_sampler_states[sampler_id].is_nonnormalized_coords = conf ? !conf->args[i].sampler.normalized_coords : 0; int_sampler_states[sampler_id].is_linear_filtering = conf ? conf->args[i].sampler.linear_filtering : 0; metadata->args[i].sampler.sampler_id = var->data.binding = sampler_id++; } } unsigned num_global_inputs = uav_id; // Second pass over inputs to calculate image bindings unsigned srv_id = 0; nir_foreach_variable_with_modes(var, nir, nir_var_uniform) { int i = var->data.location; if (i < 0) continue; if (glsl_type_is_image(var->type)) { if (var->data.access == ACCESS_NON_WRITEABLE) { metadata->args[i].image.buf_ids[0] = srv_id++; } else { // Write or read-write are UAVs metadata->args[i].image.buf_ids[0] = uav_id++; } metadata->args[i].image.num_buf_ids = 1; var->data.binding = metadata->args[i].image.buf_ids[0]; } } // Before removing dead uniforms, dedupe constant samplers to make more dead uniforms NIR_PASS_V(nir, clc_nir_dedupe_const_samplers); NIR_PASS_V(nir, nir_remove_dead_variables, nir_var_uniform | nir_var_mem_ubo | nir_var_mem_constant | nir_var_function_temp, NULL); // Fill out inline sampler metadata, now that they've been deduped and dead ones removed nir_foreach_variable_with_modes(var, nir, nir_var_uniform) { if (glsl_type_is_sampler(var->type) && var->data.sampler.is_inline_sampler) { int_sampler_states[sampler_id].wrap[0] = int_sampler_states[sampler_id].wrap[1] = int_sampler_states[sampler_id].wrap[2] = wrap_from_cl_addressing(var->data.sampler.addressing_mode); int_sampler_states[sampler_id].is_nonnormalized_coords = !var->data.sampler.normalized_coordinates; int_sampler_states[sampler_id].is_linear_filtering = var->data.sampler.filter_mode == SAMPLER_FILTER_MODE_LINEAR; var->data.binding = sampler_id++; assert(metadata->num_const_samplers < CLC_MAX_SAMPLERS); metadata->const_samplers[metadata->num_const_samplers].sampler_id = var->data.binding; metadata->const_samplers[metadata->num_const_samplers].addressing_mode = var->data.sampler.addressing_mode; metadata->const_samplers[metadata->num_const_samplers].normalized_coords = var->data.sampler.normalized_coordinates; metadata->const_samplers[metadata->num_const_samplers].filter_mode = var->data.sampler.filter_mode; metadata->num_const_samplers++; } } // Needs to come before lower_explicit_io NIR_PASS_V(nir, nir_lower_readonly_images_to_tex, false); struct clc_image_lower_context image_lower_context = { metadata, &srv_id, &uav_id }; NIR_PASS_V(nir, clc_lower_images, &image_lower_context); NIR_PASS_V(nir, clc_lower_nonnormalized_samplers, int_sampler_states); NIR_PASS_V(nir, nir_lower_samplers); NIR_PASS_V(nir, dxil_lower_sample_to_txf_for_integer_tex, int_sampler_states, NULL, 14.0f); NIR_PASS_V(nir, nir_remove_dead_variables, nir_var_mem_shared | nir_var_function_temp, NULL); nir->scratch_size = 0; NIR_PASS_V(nir, nir_lower_vars_to_explicit_types, nir_var_mem_shared | nir_var_function_temp | nir_var_mem_global | nir_var_mem_constant, glsl_get_cl_type_size_align); NIR_PASS_V(nir, dxil_nir_lower_ubo_to_temp); NIR_PASS_V(nir, clc_lower_constant_to_ssbo, out_dxil->kernel, &uav_id); NIR_PASS_V(nir, clc_lower_global_to_ssbo); bool has_printf = false; NIR_PASS(has_printf, nir, clc_lower_printf_base, uav_id); metadata->printf.uav_id = has_printf ? uav_id++ : -1; NIR_PASS_V(nir, dxil_nir_lower_deref_ssbo); NIR_PASS_V(nir, split_unaligned_loads_stores); assert(nir->info.cs.ptr_size == 64); NIR_PASS_V(nir, nir_lower_explicit_io, nir_var_mem_ssbo, nir_address_format_32bit_index_offset_pack64); NIR_PASS_V(nir, nir_lower_explicit_io, nir_var_mem_shared | nir_var_function_temp | nir_var_uniform, nir_address_format_32bit_offset_as_64bit); NIR_PASS_V(nir, nir_lower_system_values); nir_lower_compute_system_values_options compute_options = { .has_base_global_invocation_id = (conf && conf->support_global_work_id_offsets), .has_base_workgroup_id = (conf && conf->support_workgroup_id_offsets), }; NIR_PASS_V(nir, nir_lower_compute_system_values, &compute_options); NIR_PASS_V(nir, clc_lower_64bit_semantics); NIR_PASS_V(nir, nir_opt_deref); NIR_PASS_V(nir, nir_lower_vars_to_ssa); unsigned cbv_id = 0; nir_variable *inputs_var = add_kernel_inputs_var(out_dxil, nir, &cbv_id); nir_variable *work_properties_var = add_work_properties_var(out_dxil, nir, &cbv_id); memcpy(metadata->local_size, nir->info.workgroup_size, sizeof(metadata->local_size)); memcpy(metadata->local_size_hint, nir->info.cs.workgroup_size_hint, sizeof(metadata->local_size)); // Patch the localsize before calling clc_nir_lower_system_values(). if (conf) { for (unsigned i = 0; i < ARRAY_SIZE(nir->info.workgroup_size); i++) { if (!conf->local_size[i] || conf->local_size[i] == nir->info.workgroup_size[i]) continue; if (nir->info.workgroup_size[i] && nir->info.workgroup_size[i] != conf->local_size[i]) { debug_printf("D3D12: runtime local size does not match reqd_work_group_size() values\n"); goto err_free_dxil; } nir->info.workgroup_size[i] = conf->local_size[i]; } memcpy(metadata->local_size, nir->info.workgroup_size, sizeof(metadata->local_size)); } else { /* Make sure there's at least one thread that's set to run */ for (unsigned i = 0; i < ARRAY_SIZE(nir->info.workgroup_size); i++) { if (nir->info.workgroup_size[i] == 0) nir->info.workgroup_size[i] = 1; } } NIR_PASS_V(nir, clc_nir_lower_kernel_input_loads, inputs_var); NIR_PASS_V(nir, split_unaligned_loads_stores); NIR_PASS_V(nir, nir_lower_explicit_io, nir_var_mem_ubo, nir_address_format_32bit_index_offset); NIR_PASS_V(nir, clc_nir_lower_system_values, work_properties_var); NIR_PASS_V(nir, dxil_nir_lower_loads_stores_to_dxil); NIR_PASS_V(nir, dxil_nir_opt_alu_deref_srcs); NIR_PASS_V(nir, dxil_nir_lower_atomics_to_dxil); NIR_PASS_V(nir, nir_lower_fp16_casts); NIR_PASS_V(nir, nir_lower_convert_alu_types, NULL); // Convert pack to pack_split NIR_PASS_V(nir, nir_lower_pack); // Lower pack_split to bit math NIR_PASS_V(nir, nir_opt_algebraic); NIR_PASS_V(nir, nir_opt_dce); nir_validate_shader(nir, "Validate before feeding NIR to the DXIL compiler"); struct nir_to_dxil_options opts = { .interpolate_at_vertex = false, .lower_int16 = (conf && (conf->lower_bit_size & 16) != 0), .ubo_binding_offset = 0, .disable_math_refactoring = true, .num_kernel_globals = num_global_inputs, }; for (unsigned i = 0; i < out_dxil->kernel->num_args; i++) { if (out_dxil->kernel->args[i].address_qualifier != CLC_KERNEL_ARG_ADDRESS_LOCAL) continue; /* If we don't have the runtime conf yet, we just create a dummy variable. * This will be adjusted when clc_spirv_to_dxil() is called with a conf * argument. */ unsigned size = 4; if (conf && conf->args) size = conf->args[i].localptr.size; /* The alignment required for the pointee type is not easy to get from * here, so let's base our logic on the size itself. Anything bigger than * the maximum alignment constraint (which is 128 bytes, since ulong16 or * doubl16 size are the biggest base types) should be aligned on this * maximum alignment constraint. For smaller types, we use the size * itself to calculate the alignment. */ unsigned alignment = size < 128 ? (1 << (ffs(size) - 1)) : 128; nir->info.shared_size = align(nir->info.shared_size, alignment); metadata->args[i].localptr.sharedmem_offset = nir->info.shared_size; nir->info.shared_size += size; } metadata->local_mem_size = nir->info.shared_size; metadata->priv_mem_size = nir->scratch_size; /* DXIL double math is too limited compared to what NIR expects. Let's refuse * to compile a shader when it contains double operations until we have * double lowering hooked up. */ if (shader_has_double(nir)) { clc_error(logger, "NIR shader contains doubles, which we don't support yet"); goto err_free_dxil; } struct blob tmp; if (!nir_to_dxil(nir, &opts, &tmp)) { debug_printf("D3D12: nir_to_dxil failed\n"); goto err_free_dxil; } nir_foreach_variable_with_modes(var, nir, nir_var_mem_ssbo) { if (var->constant_initializer) { if (glsl_type_is_array(var->type)) { int size = align(glsl_get_cl_size(var->type), 4); uint8_t *data = malloc(size); if (!data) goto err_free_dxil; copy_const_initializer(var->constant_initializer, var->type, data); metadata->consts[metadata->num_consts].data = data; metadata->consts[metadata->num_consts].size = size; metadata->consts[metadata->num_consts].uav_id = var->data.binding; metadata->num_consts++; } else unreachable("unexpected constant initializer"); } } metadata->kernel_inputs_cbv_id = inputs_var ? inputs_var->data.binding : 0; metadata->work_properties_cbv_id = work_properties_var->data.binding; metadata->num_uavs = uav_id; metadata->num_srvs = srv_id; metadata->num_samplers = sampler_id; ralloc_free(nir); glsl_type_singleton_decref(); blob_finish_get_buffer(&tmp, &out_dxil->binary.data, &out_dxil->binary.size); return true; err_free_dxil: clc_free_dxil_object(out_dxil); return false; } void clc_free_dxil_object(struct clc_dxil_object *dxil) { for (unsigned i = 0; i < dxil->metadata.num_consts; i++) free(dxil->metadata.consts[i].data); for (unsigned i = 0; i < dxil->metadata.printf.info_count; i++) { free(dxil->metadata.printf.infos[i].arg_sizes); free(dxil->metadata.printf.infos[i].str); } free(dxil->metadata.printf.infos); free(dxil->binary.data); } uint64_t clc_compiler_get_version() { const char sha1[] = MESA_GIT_SHA1; const char* dash = strchr(sha1, '-'); if (dash) { return strtoull(dash + 1, NULL, 16); } return 0; }